Patent Application
20240402934
The technical field to which the invention relates is generally data storage systems, more specifically garbage collection, and data compression, in data storage systems.
There are many types and architectures of data storage systems, with various types of storage memory, and various schemes for data reliability and data storage efficiency. Many storage systems employ garbage collection in storage memory, in which live data is relocated in storage memory so that erase blocks with dead data can be erased and reused, also called reclaiming storage memory. Many storage systems employ various types of data compression, in which identical and/or similar data is recognized and compressed for greater data storage efficiency. Data deduplication is another strategy for data storage efficiency. As storage systems are scaled up, and as there is an ever-increasing need for storage of more data, there is an ongoing need for improvements in data storage technology, garbage collection technology, and data compression technology. It is in this environment of computing and data storage technology that present embodiments arise.
with some embodiments.
Storage systems, also referred to as data storage systems, are described with various mechanisms for data reliability and data storage efficiency with reference to
During garbage collection, and for improved data storage efficiency, data storage system embodiments described herein, and variations thereof, group blocks with similar content together and compress those blocks together to achieve improved data reduction. To do this, the storage system identifies which segments have similar blocks in them, so the storage system can run garbage collection (GC) on those segments at the same time to allow such grouping. Once the storage system has done this grouping, the storage system can compress and write out the new segments containing similar blocks, obtaining better data compression.
The storage system can describe each segment using fingerprints of the segment that record the hashes of data chunks in the segment. These chunks would likely be small (on the order of tens of bytes); since there would be a relatively large amount of them, the storage system only stores a subset of the hashes, which can be done by taking the k minimum hashes (i.e., a count of k hashes that have the lower hash values in the original full set of hashes of the segment), k maximum hashes (i.e., a count of k hashes that have the upper hash values in the original full set of hashes of the segment), or some other technique by which the storage system deterministically selects hashes (also referred to as hash values or hash results) such that independent selection processes pick the same hashes. These hashes could be stored with the segment or computed each time the storage system wants to do GC. The storage system then computes the Jaccard distance between segments, or uses some other similarity metric, and uses that as a metric to select segments to garbage collect together. The storage system in some embodiments can further compute similar metrics for individual blocks in the segments undergoing GC, and use the information to route live blocks to segments with other similar data, so that data segments with determined amounts of similarity of data according to a similarity metric are garbage collected and compressed together.
In some embodiments, this information can be used in conjunction with other information (e.g., age, percent dirty, or other data characteristics) to select segments for GC. Embodiments of storage systems can also use this information to route data in a system to a node that will “specialize” in storing blocks with particular data patterns, for example types of data, data of a specific application or client, etc.
Similarity hashes are a technique for determining that two or more blocks have similar or identical content. Several schemes for computing hashes are available in the literature, and generally present their results as a degree of distance between one block and another.
One of the most common techniques to quickly determine similarity of two blocks is to generate a large set of relatively small hashes of various sizes, and to select a subset of those hashes to represent the entire block. For example, one could perform a “rolling” hash over overlapping k-byte sequences (typically, k is between 4 and 8), generating just under 4096 hashes per 4 KiB block. It should be appreciated that in some embodiments it is important to actually hash the values (e.g., using cityhash or similar) to ensure a good distribution of values. Then, the algorithm selects the n smallest hashes as the fingerprint of the block. If this is done for two different blocks, the number of similar k byte sequences in the two blocks is about the same as the number of matching hashes in the fingerprints.
Given that two or more blocks are similar in some manner, there are often data reduction benefits to be had in grouping them together to run against a compression algorithm, as they will likely have common substrings or a similar distribution of common byte values. Even non-identical pre-compressed data (such as various blocks of already compressed data received by a storage system that were themselves compressed from similar but not identical data) can, in some instances, be further compressed if two chunks of original data were compressed using the same algorithm and produced similar compression dictionaries (in which case, at least the dictionary part of the compressed data may include common substrings).
Storage system embodiments described herein arrange for similar blocks to be gathered together when writing a large data segment, for use in storage systems that can gather blocks together from multiple sources to be written out into a shared large data segment. For example, in a storage system that writes large erasure coded stripes sharded across a set of flash storage devices, or across a set of drives that support zoned storage, it is common to gather data in large data segments with sizes on the order of megabytes, or even tens or hundreds of megabytes, to write out as sharded erasure coded stripes. It is also increasingly common for storage systems to eschew overwrites of those stored large data segments even as logical content of the storage system is replaced, and to instead use garbage collection techniques to delete content from these large data segments that are no longer needed. The remaining content from those large segments may be written to new large data segments, which are themselves gathered together from multiple sources, to preserve content from earlier large data segments that is still needed for the storage system's stored datasets.
In such a storage system that operates by collecting data from multiple sources to be stored in collective large data segments, it is possible for the mapping of logical content to be based on some combination of a reference to the large data segment and some additional data that describes how a particular smaller chunk of data can be extracted from that large data segment or block. That additional data could be a block index, or byte index, within the large data segment, or it could be a logical index within a collection of blocks that were compressed together (for example, bytes 10,000 through 16,000 of a set of a chunk of data that was then compressed before being stored in the large data segment). These techniques can be combined, where a block or byte index into a large data segment can point to a compressed block containing multiple data blocks that should be decompressed together with an index into the compressed data to locate the particular stored logical content.
Combining these ideas together, a storage system can use similarity hashes to decide which large data segments, or even which parts of a large data segment, particular blocks of data should be gathered together into, resulting in data that can compress together efficiently being brought together to be compressed and written together within the same large data segment. This algorithm can be used anytime data is being collected into some number of large data segments, which includes when data is first being written to the storage system (or after it has first been written to some first-stage durable cache and is subsequently being written to bulk durable storage) and any subsequent garbage collection passes. Garbage collection passes present the most opportunities in some embodiments, as similarity hashes themselves can drive the priority and scheduling for garbage collecting particular prior large data segments to form new ones.
It should be appreciated that a common technique for managing performance at scale in a scale-out storage system is to divide up responsibility for managing particular subsets of data at the logical level, where different processing nodes with compute and memory resources handle the interlocking, logical consistency, and other aspects of proper storage handling, of distinct subset of the data or of distinct logical structures within the logical stored data. For example, files can be divided up and assigned to processing nodes based on their inode number or some other file identifier; or, subranges of a file, or block ranges of a volume, can be divided up by range so that performance for large files can be improved by processing separate parts of the file in parallel across multiple processing nodes.
Another common feature of scale-out storage is that particular elements of physical durable storage (e.g., Flash memory) are accessible from a small subset of all processing nodes. In addition, in many cases a particular element of physical durable storage is contained within a processing node and can only be accessed by communicating to the processing node that can access (or that physically contains) that element of physical durable storage. In a scale-out storage system that writes large data segments, it can be particularly efficient to let each different node handle the accumulation of data into particular large data segments. Consequently, large data segments are erasure coded and sharded across elements of physical durable storage that are accessed through (or contained within) separate processing nodes, such that writing a large data segment requires sending data to other processing nodes in order to write the large data segment as an erasure coded stripe. Even in such a case, there are many advantages to handling control of particular logical elements within one processing node, as it greatly reduces the complexity of locking and ensuring consistency for operations against each particular logical element.
Bringing separate chunks of data together by their similarity hash, though, introduces another type of coordination that may be thought of as being orthogonal to division by logical element. Particular chunks of data that are similar, and should be written together as combined compressed blocks, do not follow a pattern that can be hashed to processing nodes by logical elements such as file identifier or block address, since similarity hashes are derived from content and not by how that content is logically addressed. It should be appreciated that this feature gives rise to opportunities to address possible issues. There are now two layers which operate across processing nodes that have to be coordinated. One layer is the layer that handles coordination of logical structure, which may be handled by a logical element processing node. The other layer determines which data chunks are similar and arranges for similar data chunks to be compressed together, which for example is handled by a similarity compression processing node, all before the results of that compression can then be sharded across storage element processing nodes to be written as an erasure coded stripe to separate elements of physical durable storage (all these types of processing nodes will usually be different aspects of compute logic running on the same sets of processing nodes). The data to be compressed together should reside in the same memory on the same similarity compression processing node to be run through the compression algorithm together, all before eventually being transferred to the various other storage element processing nodes to be written out as the shards of the large erasure coded data segment stripe. In addition, the amount of compression, and thus the amount of data that can be included in a particular large data segment, cannot be known at the time a logical element processing node handling content for a particular logical element is allocating writes for that content into large data segments. Thus, when processing a particular logical element, the logic doing that processing cannot really know which large data segment it will end up in.
Some or all of these opportunities can be mitigated by decoupling the time at which content is gathered together by similarity hash, compressed, and accumulated into large data segments from the time an operation against the storage system affects a logical element. A garbage collection pass, for example, can be mostly decoupled from operations against logical elements, with the only interaction being the updating of any pointers to where particular data is stored, a task that is often needed when doing garbage collection anyway. If similarity hashes are calculated as data comes into a storage system or as the data is being initially stored, then in-memory or durable indexes can be updated so that a garbage collection pass can know at the time of the garbage collection pass which data chunks are similar and benefit from being compressed together. A similarity compression processing node can read chunks from a variety of large data segments that are matched together by similarity hash and compress them together. The processing node can continue accumulating additional chunks of data with additional similarity hashes that can then be separately compressed together, accumulating and writing out the resulting compressed data into new large data segments to be written out through storage element processing nodes. References to the data chunks can then be updated, perhaps through logical element processing nodes, and remaining chunks within prior large data segments can then be considered garbage when those prior large data segments are finally garbage collected, with their remaining data written elsewhere before eventually being erased and reused.
It should be appreciated that the above processing steps have minimal interaction with logical elements, and no more interaction than a normal garbage collection pass. The difference is that the processing nodes involved are divided up in a way that data with the same similarity hash is processed in any garbage collection pass by the same similarity compression processing node. Any number of such processing nodes can operate concurrently against distinct sets of similarity hashes, similar to common implementations of scale-out storage large data segment garbage collection that don't operate using similarity hashes.
Implementing this similarity hash compression model can present a combinatorial problem, where an existing large data segment can include data chunks that match similarity hashes in many other large data segments, and where those other large data segments include data chunks that match similarity hashes in many other large data segments, and so on. As a result, given a stored dataset that has not already compressed together data chunks with matching similarity hashes, particularly a dataset with a large number of randomly scattered matching similarity hashes, it could be quite expensive computationally to bring them together as part of garbage collection. In some embodiments it may make more sense to run the two processes separately, with one process identifying and bringing together similar blocks, resulting in the blocks at their original locations becoming “garbage” that is eventually garbage collected by the standard garbage collector.
The same algorithm can be employed when reprocessing newly arrived data after it has been stored in first-stage durable cache. As the data is being transformed in the process of writing to bulk durable storage, that data can be merged with the process described above for rewriting existing data into large data segments based on matching similarity hashes in some embodiments.
It should be further appreciated that in storage systems described herein with authorities the authorities handle the management of segments through their entire lifespan. This model assumes that data from any sources can be brought together to be compressed together into the same segment, which in turn means that one implementation reduces data within an authority. Some storage system embodiments could have a separation between segment ownership and higher-level structure, which would then require a different messaging model between blades, and would affect the logic of how authorities (and garbage collection) work.
Some embodiments may sort segments over time by looking at groups of segments that are garbage collection candidates and then sorting blocks within those segments by similarity distance, producing benefit without requiring a global table. Some embodiments may look for better candidate blocks over time through random sampling that can improve the odds of identifying segments that should be garbage collected together, at the same time. With authorities, this sorting could be implemented within individual authorities, improving data reduction within an authority managed dataset, but not necessarily to the degree that might be possible with a more global implementation.
Some embodiments find common base blocks and build them as a dictionary (which may be used as a basis for differential compression). This embodiment is suitable for storage systems that have authorities, as it is likely to define a relatively limited set of immutable base blocks that authorities can store differences to on an authority-by-authority basis. Alternatively, some embodiments generalize the process to run a full cross-authority model for compressing similar blocks together. Some embodiments may employ direct mapped flash memory to simplify forward looking feature development by moving away from the complexity of pyramids. In some embodiments a mix of drive types, in various drive slots may be utilized in the system. These embodiments may utilize side tables in a more random form of storage, such as 3D cross point.
Example methods, apparatus, and products for data storage systems in accordance with embodiments of the present disclosure are described with reference to the accompanying drawings, beginning with
System 100 includes a number of computing devices 164A-B. Computing devices (also referred to as “client devices” herein) may be embodied, for example, a server in a data center, a workstation, a personal computer, a notebook, or the like. Computing devices 164A-B may be coupled for data communications to one or more storage arrays 102A-B through a storage area network (‘SAN’) 158 or a local area network (‘LAN’) 160.
The SAN 158 may be implemented with a variety of data communications fabrics, devices, and protocols. For example, the fabrics for SAN 158 may include Fibre Channel, Ethernet, Infiniband, Serial Attached Small Computer System Interface (‘SAS’), or the like. Data communications protocols for use with SAN 158 may include Advanced Technology Attachment (‘ATA’), Fibre Channel Protocol, Small Computer System Interface (‘SCSI’), Internet Small Computer System Interface (‘iSCSI’), HyperSCSI, Non-Volatile Memory Express (‘NVMe’) over Fabrics, or the like. It may be noted that SAN 158 is provided for illustration, rather than limitation. Other data communication couplings may be implemented between computing devices 164A-B and storage arrays 102A-B.
The LAN 160 may also be implemented with a variety of fabrics, devices, and protocols. For example, the fabrics for LAN 160 may include Ethernet (802.3), wireless (802.11), or the like. Data communication protocols for use in LAN 160 may include Transmission Control Protocol (‘TCP’), User Datagram Protocol (‘UDP’), Internet Protocol (‘IP’), HyperText Transfer Protocol (‘HTTP’), Wireless Access Protocol (‘WAP’), Handheld Device Transport Protocol (‘HDTP’), Session Initiation Protocol (‘SIP’), Real Time Protocol (‘RTP’), or the like.
Storage arrays 102A-B may provide persistent data storage for the computing devices 164A-B. Storage array 102A may be contained in a chassis (not shown), and storage array 102B may be contained in another chassis (not shown), in implementations. Storage array 102A and 102B may include one or more storage array controllers 110A-D (also referred to as “controller” herein). A storage array controller 110A-D may be embodied as a module of automated computing machinery comprising computer hardware, computer software, or a combination of computer hardware and software. In some implementations, the storage array controllers 110A-D may be configured to carry out various storage tasks. Storage tasks may include writing data received from the computing devices 164A-B to storage array 102A-B, erasing data from storage array 102A-B, retrieving data from storage array 102A-B and providing data to computing devices 164A-B, monitoring and reporting of disk utilization and performance, performing redundancy operations, such as Redundant Array of Independent Drives (‘RAID’) or RAID-like data redundancy operations, compressing data, encrypting data, and so forth.
Storage array controller 110A-D may be implemented in a variety of ways, including as a Field Programmable Gate Array (‘FPGA’), a Programmable Logic Chip (‘PLC’), an Application Specific Integrated Circuit (‘ASIC’), System-on-Chip (‘SOC’), or any computing device that includes discrete components such as a processing device, central processing unit, computer memory, or various adapters. Storage array controller 110A-D may include, for example, a data communications adapter configured to support communications via the SAN 158 or LAN 160. In some implementations, storage array controller 110A-D may be independently coupled to the LAN 160. In implementations, storage array controller 110A-D may include an I/O controller or the like that couples the storage array controller 110A-D for data communications, through a midplane (not shown), to a persistent storage resource 170A-B (also referred to as a “storage resource” herein). The persistent storage resource 170A-B main include any number of storage drives 171A-F (also referred to as “storage devices” herein) and any number of non-volatile Random Access Memory (‘NVRAM’) devices (not shown).
In some implementations, the NVRAM devices of a persistent storage resource 170A-B may be configured to receive, from the storage array controller 110A-D, data to be stored in the storage drives 171A-F. In some examples, the data may originate from computing devices 164A-B. In some examples, writing data to the NVRAM device may be carried out more quickly than directly writing data to the storage drive 171A-F. In implementations, the storage array controller 110A-D may be configured to utilize the NVRAM devices as a quickly accessible buffer for data destined to be written to the storage drives 171A-F. Latency for write requests using NVRAM devices as a buffer may be improved relative to a system in which a storage array controller 110A-D writes data directly to the storage drives 171A-F. In some implementations, the NVRAM devices may be implemented with computer memory in the form of high bandwidth, low latency RAM. The NVRAM device is referred to as “non-volatile” because the NVRAM device may receive or include a unique power source that maintains the state of the RAM after main power loss to the NVRAM device. Such a power source may be a battery, one or more capacitors, or the like. In response to a power loss, the NVRAM device may be configured to write the contents of the RAM to a persistent storage, such as the storage drives 171A-F.
In implementations, storage drive 171A-F may refer to any device configured to record data persistently, where “persistently” or “persistent” refers as to a device's ability to maintain recorded data after loss of power. In some implementations, storage drive 171A-F may correspond to non-disk storage media. For example, the storage drive 171A-F may be one or more solid-state drives (‘SSDs’), flash memory based storage, any type of solid-state non-volatile memory, or any other type of non-mechanical storage device. In other implementations, storage drive 171A-F may include mechanical or spinning hard disk, such as hard-disk drives ('HDD').
In some implementations, the storage array controllers 110A-D may be configured for offloading device management responsibilities from storage drive 171A-F in storage array 102A-B. For example, storage array controllers 110A-D may manage control information that may describe the state of one or more memory blocks in the storage drives 171A-F. The control information may indicate, for example, that a particular memory block has failed and should no longer be written to, that a particular memory block contains boot code for a storage array controller 110A-D, the number of program-erase (‘P/E’) cycles that have been performed on a particular memory block, the age of data stored in a particular memory block, the type of data that is stored in a particular memory block, and so forth. In some implementations, the control information may be stored with an associated memory block as metadata. In other implementations, the control information for the storage drives 171A-F may be stored in one or more particular memory blocks of the storage drives 171A-F that are selected by the storage array controller 110A-D. The selected memory blocks may be tagged with an identifier indicating that the selected memory block contains control information. The identifier may be utilized by the storage array controllers 110A-D in conjunction with storage drives 171A-F to quickly identify the memory blocks that contain control information. For example, the storage controllers 110A-D may issue a command to locate memory blocks that contain control information. It may be noted that control information may be so large that parts of the control information may be stored in multiple locations, that the control information may be stored in multiple locations for purposes of redundancy, for example, or that the control information may otherwise be distributed across multiple memory blocks in the storage drive 171A-F.
In implementations, storage array controllers 110A-D may offload device management responsibilities from storage drives 171A-F of storage array 102A-B by retrieving, from the storage drives 171A-F, control information describing the state of one or more memory blocks in the storage drives 171A-F. Retrieving the control information from the storage drives 171A-F may be carried out, for example, by the storage array controller 110A-D querying the storage drives 171A-F for the location of control information for a particular storage drive 171A-F. The storage drives 171A-F may be configured to execute instructions that enable the storage drive 171A-F to identify the location of the control information. The instructions may be executed by a controller (not shown) associated with or otherwise located on the storage drive 171A-F and may cause the storage drive 171A-F to scan a portion of each memory block to identify the memory blocks that store control information for the storage drives 171A-F. The storage drives 171A-F may respond by sending a response message to the storage array controller 110A-D that includes the location of control information for the storage drive 171A-F. Responsive to receiving the response message, storage array controllers 110A-D may issue a request to read data stored at the address associated with the location of control information for the storage drives 171A-F.
In other implementations, the storage array controllers 110A-D may further offload device management responsibilities from storage drives 171A-F by performing, in response to receiving the control information, a storage drive management operation. A storage drive management operation may include, for example, an operation that is typically performed by the storage drive 171A-F (e.g., the controller (not shown) associated with a particular storage drive 171A-F). A storage drive management operation may include, for example, ensuring that data is not written to failed memory blocks within the storage drive 171A-F, ensuring that data is written to memory blocks within the storage drive 171A-F in such a way that adequate wear leveling is achieved, and so forth.
In implementations, storage array 102A-B may implement two or more storage array controllers 110A-D. For example, storage array 102A may include storage array controllers 110A and storage array controllers 110B. At a given instance, a single storage array controller 110A-D (e.g., storage array controller 110A) of a storage system 100 may be designated with primary status (also referred to as “primary controller” herein), and other storage array controllers 110A-D (e.g., storage array controller 110A) may be designated with secondary status (also referred to as “secondary controller” herein). The primary controller may have particular rights, such as permission to alter data in persistent storage resource 170A-B (e.g., writing data to persistent storage resource 170A-B). At least some of the rights of the primary controller may supersede the rights of the secondary controller. For instance, the secondary controller may not have permission to alter data in persistent storage resource 170A-B when the primary controller has the right. The status of storage array controllers 110A-D may change. For example, storage array controller 110A may be designated with secondary status, and storage array controller 110B may be designated with primary status.
In some implementations, a primary controller, such as storage array controller 110A, may serve as the primary controller for one or more storage arrays 102A-B, and a second controller, such as storage array controller 110B, may serve as the secondary controller for the one or more storage arrays 102A-B. For example, storage array controller 110A may be the primary controller for storage array 102A and storage array 102B, and storage array controller 110B may be the secondary controller for storage array 102A and 102B. In some implementations, storage array controllers 110C and 110D (also referred to as “storage processing modules”) may neither have primary or secondary status. Storage array controllers 110C and 110D, implemented as storage processing modules, may act as a communication interface between the primary and secondary controllers (e.g., storage array controllers 110A and 110B, respectively) and storage array 102B. For example, storage array controller 110A of storage array 102A may send a write request, via SAN 158, to storage array 102B. The write request may be received by both storage array controllers 110C and 110D of storage array 102B. Storage array controllers 110C and 110D facilitate the communication, e.g., send the write request to the appropriate storage drive 171A-F. It may be noted that in some implementations storage processing modules may be used to increase the number of storage drives controlled by the primary and secondary controllers.
In implementations, storage array controllers 110A-D are communicatively coupled, via a midplane (not shown), to one or more storage drives 171A-F and to one or more NVRAM devices (not shown) that are included as part of a storage array 102A-B. The storage array controllers 110A-D may be coupled to the midplane via one or more data communication links and the midplane may be coupled to the storage drives 171A-F and the NVRAM devices via one or more data communications links. The data communications links described herein are collectively illustrated by data communications links 108A-D and may include a Peripheral Component Interconnect Express (‘PCIe’) bus, for example.
Storage array controller 101 may include one or more processing devices 104 and random access memory (‘RAM’) 111. Processing device 104 (or controller 101) represents one or more general-purpose processing devices such as a microprocessor, central processing unit, or the like. More particularly, the processing device 104 (or controller 101) may be a complex instruction set computing (‘CISC’) microprocessor, reduced instruction set computing (‘RISC’) microprocessor, very long instruction word (‘VLIW’) microprocessor, or a processor implementing other instruction sets or processors implementing a combination of instruction sets. The processing device 104 (or controller 101) may also be one or more special-purpose processing devices such as an ASIC, an FPGA, a digital signal processor (‘DSP’), network processor, or the like.
The processing device 104 may be connected to the RAM 111 via a data communications link 106, which may be embodied as a high speed memory bus such as a Double-Data Rate 4(‘DDR4’) bus. Stored in RAM 111 is an operating system 112. In some implementations, instructions 113 are stored in RAM 111. Instructions 113 may include computer program instructions for performing operations in in a direct-mapped flash storage system. In one embodiment, a direct-mapped flash storage system is one that addresses data blocks within flash drives directly and without an address translation performed by the storage controllers of the flash drives.
In implementations, storage array controller 101 includes one or more host bus adapters 103A-C that are coupled to the processing device 104 via a data communications link 105A-C. In implementations, host bus adapters 103A-C may be computer hardware that connects a host system (e.g., the storage array controller) to other network and storage arrays. In some examples, host bus adapters 103A-C may be a Fibre Channel adapter that enables the storage array controller 101 to connect to a SAN, an Ethernet adapter that enables the storage array controller 101 to connect to a LAN, or the like. Host bus adapters 103A-C may be coupled to the processing device 104 via a data communications link 105A-C such as, for example, a PCIe bus.
In implementations, storage array controller 101 may include a host bus adapter 114 that is coupled to an expander 115. The expander 115 may be used to attach a host system to a larger number of storage drives. The expander 115 may, for example, be a SAS expander utilized to enable the host bus adapter 114 to attach to storage drives in an implementation where the host bus adapter 114 is embodied as a SAS controller.
In implementations, storage array controller 101 may include a switch 116 coupled to the processing device 104 via a data communications link 109. The switch 116 may be a computer hardware device that can create multiple endpoints out of a single endpoint, thereby enabling multiple devices to share a single endpoint. The switch 116 may, for example, be a PCIe switch that is coupled to a PCIe bus (e.g., data communications link 109) and presents multiple PCIe connection points to the midplane.
In implementations, storage array controller 101 includes a data communications link 107 for coupling the storage array controller 101 to other storage array controllers. In some examples, data communications link 107 may be a QuickPath Interconnect (QPI) interconnect.
A traditional storage system that uses traditional flash drives may implement a process across the flash drives that are part of the traditional storage system. For example, a higher level process of the storage system may initiate and control a process across the flash drives. However, a flash drive of the traditional storage system may include its own storage controller that also performs the process. Thus, for the traditional storage system, a higher level process (e.g., initiated by the storage system) and a lower level process (e.g., initiated by a storage controller of the storage system) may both be performed.
To resolve various deficiencies of a traditional storage system, operations may be performed by higher level processes and not by the lower level processes. For example, the flash storage system may include flash drives that do not include storage controllers that provide the process. Thus, the operating system of the flash storage system itself may initiate and control the process. This may be accomplished by a direct-mapped flash storage system that addresses data blocks within the flash drives directly and without an address translation performed by the storage controllers of the flash drives.
In implementations, storage drive 171A-F may be one or more zoned storage devices. In some implementations, the one or more zoned storage devices may be a shingled HDD. In implementations, the one or more storage devices may be a flash-based SSD. In a zoned storage device, a zoned namespace on the zoned storage device can be addressed by groups of blocks that are grouped and aligned by a natural size, forming a number of addressable zones. In implementations utilizing an SSD, the natural size may be based on the erase block size of the SSD.
The mapping from a zone to an erase block (or to a shingled track in an HDD) may be arbitrary, dynamic, and hidden from view. The process of opening a zone may be an operation that allows a new zone to be dynamically mapped to underlying storage of the zoned storage device, and then allows data to be written through appending writes into the zone until the zone reaches capacity. The zone can be finished at any point, after which further data may not be written into the zone. When the data stored at the zone is no longer needed, the zone can be reset which effectively deletes the zone's content from the zoned storage device, making the physical storage held by that zone available for the subsequent storage of data. Once a zone has been written and finished, the zoned storage device ensures that the data stored at the zone is not lost until the zone is reset. In the time between writing the data to the zone and the resetting of the zone, the zone may be moved around between shingle tracks or erase blocks as part of maintenance operations within the zoned storage device, such as by copying data to keep the data refreshed or to handle memory cell aging in an SSD.
In implementations utilizing an HDD, the resetting of the zone may allow the shingle tracks to be allocated to a new, opened zone that may be opened at some point in the future. In implementations utilizing an SSD, the resetting of the zone may cause the associated physical erase block(s) of the zone to be erased and subsequently reused for the storage of data. In some implementations, the zoned storage device may have a limit on the number of open zones at a point in time to reduce the amount of overhead dedicated to keeping zones open.
The operating system of the flash storage system may identify and maintain a list of allocation units across multiple flash drives of the flash storage system. The allocation units may be entire erase blocks or multiple erase blocks. The operating system may maintain a map or address range that directly maps addresses to erase blocks of the flash drives of the flash storage system.
Direct mapping to the erase blocks of the flash drives may be used to rewrite data and erase data. For example, the operations may be performed on one or more allocation units that include a first data and a second data where the first data is to be retained and the second data is no longer being used by the flash storage system. The operating system may initiate the process to write the first data to new locations within other allocation units and erasing the second data and marking the allocation units as being available for use for subsequent data. Thus, the process may only be performed by the higher level operating system of the flash storage system without an additional lower level process being performed by controllers of the flash drives.
Advantages of the process being performed only by the operating system of the flash storage system include increased reliability of the flash drives of the flash storage system as unnecessary or redundant write operations are not being performed during the process. One possible point of novelty here is the concept of initiating and controlling the process at the operating system of the flash storage system. In addition, the process can be controlled by the operating system across multiple flash drives. This is contrast to the process being performed by a storage controller of a flash drive.
A storage system can consist of two storage array controllers that share a set of drives for failover purposes, or it could consist of a single storage array controller that provides a storage service that utilizes multiple drives, or it could consist of a distributed network of storage array controllers each with some number of drives or some amount of Flash storage where the storage array controllers in the network collaborate to provide a complete storage service and collaborate on various aspects of a storage service including storage allocation and garbage collection.
In one embodiment, system 117 includes a dual Peripheral Component Interconnect (‘PCI’) flash storage device 118 with separately addressable fast write storage. System 117 may include a storage controller 119. In one embodiment, storage controller 119A-D may be a CPU, ASIC, FPGA, or any other circuitry that may implement control structures necessary according to the present disclosure. In one embodiment, system 117 includes flash memory devices (e.g., including flash memory devices 120a-n), operatively coupled to various channels of the storage device controller 119. Flash memory devices 120a-n, may be presented to the controller 119A-D as an addressable collection of Flash pages, erase blocks, and/or control elements sufficient to allow the storage device controller 119A-D to program and retrieve various aspects of the Flash. In one embodiment, storage device controller 119A-D may perform operations on flash memory devices 120a-n including storing and retrieving data content of pages, arranging and erasing any blocks, tracking statistics related to the use and reuse of Flash memory pages, erase blocks, and cells, tracking and predicting error codes and faults within the Flash memory, controlling voltage levels associated with programming and retrieving contents of Flash cells, etc.
In one embodiment, system 117 may include RAM 121 to store separately addressable fast-write data. In one embodiment, RAM 121 may be one or more separate discrete devices. In another embodiment, RAM 121 may be integrated into storage device controller 119A-D or multiple storage device controllers. The RAM 121 may be utilized for other purposes as well, such as temporary program memory for a processing device (e.g., a CPU) in the storage device controller 119.
In one embodiment, system 117 may include a stored energy device 122, such as a rechargeable battery or a capacitor. Stored energy device 122 may store energy sufficient to power the storage device controller 119, some amount of the RAM (e.g., RAM 121), and some amount of Flash memory (e.g., Flash memory 120a-120n) for sufficient time to write the contents of RAM to Flash memory. In one embodiment, storage device controller 119A-D may write the contents of RAM to Flash Memory if the storage device controller detects loss of external power.
In one embodiment, system 117 includes two data communications links 123a, 123b. In one embodiment, data communications links 123a, 123b may be PCI interfaces. In another embodiment, data communications links 123a, 123b may be based on other communications standards (e.g., HyperTransport, InfiniBand, etc.). Data communications links 123a, 123b may be based on non-volatile memory express (‘NVMe’) or NVMe over fabrics (‘NVMf’) specifications that allow external connection to the storage device controller 119A-D from other components in the storage system 117. It should be noted that data communications links may be interchangeably referred to herein as PCI buses for convenience.
System 117 may also include an external power source (not shown), which may be provided over one or both data communications links 123a, 123b, or which may be provided separately. An alternative embodiment includes a separate Flash memory (not shown) dedicated for use in storing the content of RAM 121. The storage device controller 119A-D may present a logical device over a PCI bus which may include an addressable fast-write logical device, or a distinct part of the logical address space of the storage device 118, which may be presented as PCI memory or as persistent storage. In one embodiment, operations to store into the device are directed into the RAM 121. On power failure, the storage device controller 119A-D may write stored content associated with the addressable fast-write logical storage to Flash memory (e.g., Flash memory 120a-n) for long-term persistent storage.
In one embodiment, the logical device may include some presentation of some or all of the content of the Flash memory devices 120a-n, where that presentation allows a storage system including a storage device 118 (e.g., storage system 117) to directly address Flash memory pages and directly reprogram erase blocks from storage system components that are external to the storage device through the PCI bus. The presentation may also allow one or more of the external components to control and retrieve other aspects of the Flash memory including some or all of: tracking statistics related to use and reuse of Flash memory pages, erase blocks, and cells across all the Flash memory devices; tracking and predicting error codes and faults within and across the Flash memory devices; controlling voltage levels associated with programming and retrieving contents of Flash cells; etc.
In one embodiment, the stored energy device 122 may be sufficient to ensure completion of in-progress operations to the Flash memory devices 120a-120n stored energy device 122 may power storage device controller 119A-D and associated Flash memory devices (e.g., 120a-n) for those operations, as well as for the storing of fast-write RAM to Flash memory. Stored energy device 122 may be used to store accumulated statistics and other parameters kept and tracked by the Flash memory devices 120a-n and/or the storage device controller 119. Separate capacitors or stored energy devices (such as smaller capacitors near or embedded within the Flash memory devices themselves) may be used for some or all of the operations described herein.
Various schemes may be used to track and optimize the life span of the stored energy component, such as adjusting voltage levels over time, partially discharging the storage energy device 122 to measure corresponding discharge characteristics, etc. If the available energy decreases over time, the effective available capacity of the addressable fast-write storage may be decreased to ensure that it can be written safely based on the currently available stored energy.
In one embodiment, two storage controllers (e.g., 125a and 125b) provide storage services, such as a SCS) block storage array, a file server, an object server, a database or data analytics service, etc. The storage controllers 125a, 125b may provide services through some number of network interfaces (e.g., 126a-d) to host computers 127a-n outside of the storage system 124. Storage controllers 125a, 125b may provide integrated services or an application entirely within the storage system 124, forming a converged storage and compute system. The storage controllers 125a, 125b may utilize the fast write memory within or across storage devices 119a-d to journal in progress operations to ensure the operations are not lost on a power failure, storage controller removal, storage controller or storage system shutdown, or some fault of one or more software or hardware components within the storage system 124.
In one embodiment, controllers 125a, 125b operate as PCI masters to one or the other PCI buses 128a, 128b. In another embodiment, 128a and 128b may be based on other communications standards (e.g., HyperTransport, InfiniBand, etc.). Other storage system embodiments may operate storage controllers 125a, 125b as multi-masters for both PCI buses 128a, 128b. Alternately, a PCI/NVMe/NVMf switching infrastructure or fabric may connect multiple storage controllers. Some storage system embodiments may allow storage devices to communicate with each other directly rather than communicating only with storage controllers. In one embodiment, a storage device controller 119a may be operable under direction from a storage controller 125a to synthesize and transfer data to be stored into Flash memory devices from data that has been stored in RAM (e.g., RAM 121 of
In one embodiment, under direction from a storage controller 125a, 125b, a storage device controller 119a, 119b may be operable to calculate and transfer data to other storage devices from data stored in RAM (e.g., RAM 121 of
A storage device controller 119A-D may include mechanisms for implementing high availability primitives for use by other parts of a storage system external to the Dual PCI storage device 118. For example, reservation or exclusion primitives may be provided so that, in a storage system with two storage controllers providing a highly available storage service, one storage controller may prevent the other storage controller from accessing or continuing to access the storage device. This could be used, for example, in cases where one controller detects that the other controller is not functioning properly or where the interconnect between the two storage controllers may itself not be functioning properly.
In one embodiment, a storage system for use with Dual PCI direct mapped storage devices with separately addressable fast write storage includes systems that manage erase blocks or groups of erase blocks as allocation units for storing data on behalf of the storage service, or for storing metadata (e.g., indexes, logs, etc.) associated with the storage service, or for proper management of the storage system itself. Flash pages, which may be a few kilobytes in size, may be written as data arrives or as the storage system is to persist data for long intervals of time (e.g., above a defined threshold of time). To commit data more quickly, or to reduce the number of writes to the Flash memory devices, the storage controllers may first write data into the separately addressable fast write storage on one or more storage devices.
In one embodiment, the storage controllers 125a, 125b may initiate the use of erase blocks within and across storage devices (e.g., 118) in accordance with an age and expected remaining lifespan of the storage devices, or based on other statistics. The storage controllers 125a, 125b may initiate garbage collection and data migration between storage devices in accordance with pages that are no longer needed as well as to manage Flash page and erase block lifespans and to manage overall system performance.
In one embodiment, the storage system 124 may utilize mirroring and/or erasure coding schemes as part of storing data into addressable fast write storage and/or as part of writing data into allocation units associated with erase blocks. Erasure codes may be used across storage devices, as well as within erase blocks or allocation units, or within and across Flash memory devices on a single storage device, to provide redundancy against single or multiple storage device failures or to protect against internal corruptions of Flash memory pages resulting from Flash memory operations or from degradation of Flash memory cells. Mirroring and erasure coding at various levels may be used to recover from multiple types of failures that occur separately or in combination.
The embodiments depicted with reference to
The storage cluster may be contained within a chassis, i.e., an enclosure housing one or more storage nodes. A mechanism to provide power to each storage node, such as a power distribution bus, and a communication mechanism, such as a communication bus that enables communication between the storage nodes are included within the chassis. The storage cluster can run as an independent system in one location according to some embodiments. In one embodiment, a chassis contains at least two instances of both the power distribution and the communication bus which may be enabled or disabled independently. The internal communication bus may be an Ethernet bus, however, other technologies such as PCIe, InfiniBand, and others, are equally suitable. The chassis provides a port for an external communication bus for enabling communication between multiple chassis, directly or through a switch, and with client systems. The external communication may use a technology such as Ethernet, InfiniBand, Fibre Channel, etc. In some embodiments, the external communication bus uses different communication bus technologies for inter-chassis and client communication. If a switch is deployed within or between chassis, the switch may act as a translation between multiple protocols or technologies. When multiple chassis are connected to define a storage cluster, the storage cluster may be accessed by a client using either proprietary interfaces or standard interfaces such as network file system (‘NFS’), common internet file system (‘CIFS’), small computer system interface (‘SCSI’) or hypertext transfer protocol (‘HTTP’). Translation from the client protocol may occur at the switch, chassis external communication bus or within each storage node. In some embodiments, multiple chassis may be coupled or connected to each other through an aggregator switch. A portion and/or all of the coupled or connected chassis may be designated as a storage cluster. As discussed above, each chassis can have multiple blades, each blade has a media access control (‘MAC’) address, but the storage cluster is presented to an external network as having a single cluster IP address and a single MAC address in some embodiments.
Each storage node may be one or more storage servers and each storage server is connected to one or more non-volatile solid state memory units, which may be referred to as storage units or storage devices. One embodiment includes a single storage server in each storage node and between one to eight non-volatile solid state memory units, however this one example is not meant to be limiting. The storage server may include a processor, DRAM and interfaces for the internal communication bus and power distribution for each of the power buses. Inside the storage node, the interfaces and storage unit share a communication bus, e.g., PCI Express, in some embodiments. The non-volatile solid state memory units may directly access the internal communication bus interface through a storage node communication bus, or request the storage node to access the bus interface. The non-volatile solid state memory unit contains an embedded CPU, solid state storage controller, and a quantity of solid state mass storage, e.g., between 2-32terabytes (‘TB’) in some embodiments. An embedded volatile storage medium, such as DRAM, and an energy reserve apparatus are included in the non-volatile solid state memory unit. In some embodiments, the energy reserve apparatus is a capacitor, super-capacitor, or battery that enables transferring a subset of DRAM contents to a stable storage medium in the case of power loss. In some embodiments, the non-volatile solid state memory unit is constructed with a storage class memory, such as phase change or magnetoresistive random access memory (‘MRAM’) that substitutes for DRAM and enables a reduced power hold-up apparatus.
One of many features of the storage nodes and non-volatile solid state storage is the ability to proactively rebuild data in a storage cluster. The storage nodes and non-volatile solid state storage can determine when a storage node or non-volatile solid state storage in the storage cluster is unreachable, independent of whether there is an attempt to read data involving that storage node or non-volatile solid state storage. The storage nodes and non-volatile solid state storage then cooperate to recover and rebuild the data in at least partially new locations. This constitutes a proactive rebuild, in that the system rebuilds data without waiting until the data is needed for a read access initiated from a client system employing the storage cluster. These and further details of the storage memory and operation thereof are discussed below.
Each storage node 150 can have multiple components. In the embodiment shown here, the storage node 150 includes a printed circuit board 159 populated by a CPU 156, i.e., processor, a memory 154 coupled to the CPU 156, and a non-volatile solid state storage 152 coupled to the CPU 156, although other mountings and/or components could be used in further embodiments. The memory 154 has instructions which are executed by the CPU 156 and/or data operated on by the CPU 156. As further explained below, the non-volatile solid state storage 152 includes flash or, in further embodiments, other types of solid-state memory.
Referring to
Every piece of data, and every piece of metadata, has redundancy in the system in some embodiments. In addition, every piece of data and every piece of metadata has an owner, which may be referred to as an authority. If that authority is unreachable, for example through failure of a storage node, there is a plan of succession for how to find that data or that metadata. In various embodiments, there are redundant copies of authorities 168. Authorities 168 have a relationship to storage nodes 150 and non-volatile solid state storage 152 in some embodiments. Each authority 168, covering a range of data segment numbers or other identifiers of the data, may be assigned to a specific non-volatile solid state storage 152. In some embodiments the authorities 168 for all of such ranges are distributed over the non-volatile solid state storages 152 of a storage cluster. Each storage node 150 has a network port that provides access to the non-volatile solid state storage(s) 152 of that storage node 150. Data can be stored in a segment, which is associated with a segment number and that segment number is an indirection for a configuration of a RAID (redundant array of independent disks) stripe in some embodiments. The assignment and use of the authorities 168 thus establishes an indirection to data. Indirection may be referred to as the ability to reference data indirectly, in this case via an authority 168, in accordance with some embodiments. A segment identifies a set of non-volatile solid state storage 152 and a local identifier into the set of non-volatile solid state storage 152 that may contain data. In some embodiments, the local identifier is an offset into the device and may be reused sequentially by multiple segments. In other embodiments the local identifier is unique for a specific segment and never reused. The offsets in the non-volatile solid state storage 152 are applied to locating data for writing to or reading from the non-volatile solid state storage 152 (in the form of a RAID stripe). Data is striped across multiple units of non-volatile solid state storage 152, which may include or be different from the non-volatile solid state storage 152 having the authority 168 for a particular data segment.
If there is a change in where a particular segment of data is located, e.g., during a data move or a data reconstruction, the authority 168 for that data segment should be consulted, at that non-volatile solid state storage 152 or storage node 150 having that authority 168. In order to locate a particular piece of data, embodiments calculate a hash value for a data segment or apply an inode number or a data segment number. The output of this operation points to a non-volatile solid state storage 152 having the authority 168 for that particular piece of data. In some embodiments there are two stages to this operation. The first stage maps an entity identifier (ID), e.g., a segment number, inode number, or directory number to an authority identifier. This mapping may include a calculation such as a hash or a bit mask. The second stage is mapping the authority identifier to a particular non-volatile solid state storage 152, which may be done through an explicit mapping. The operation is repeatable, so that when the calculation is performed, the result of the calculation repeatably and reliably points to a particular non-volatile solid state storage 152 having that authority 168. The operation may include the set of reachable storage nodes as input. If the set of reachable non-volatile solid state storage units changes the optimal set changes. In some embodiments, the persisted value is the current assignment (which is always true) and the calculated value is the target assignment the cluster will attempt to reconfigure towards. This calculation may be used to determine the optimal non-volatile solid state storage 152 for an authority in the presence of a set of non-volatile solid state storage 152 that are reachable and constitute the same cluster. The calculation also determines an ordered set of peer non-volatile solid state storage 152 that will also record the authority to non-volatile solid state storage mapping so that the authority may be determined even if the assigned non-volatile solid state storage is unreachable. A duplicate or substitute authority 168 may be consulted if a specific authority 168 is unavailable in some embodiments.
With reference to
In embodiments, authorities 168 operate to determine how operations will proceed against particular logical elements. Each of the logical elements may be operated on through a particular authority across a plurality of storage controllers of a storage system. The authorities 168 may communicate with the plurality of storage controllers so that the plurality of storage controllers collectively perform operations against those particular logical elements.
In embodiments, logical elements could be, for example, files, directories, object buckets, individual objects, delineated parts of files or objects, other forms of key-value pair databases, or tables. In embodiments, performing an operation can involve, for example, ensuring consistency, structural integrity, and/or recoverability with other operations against the same logical element, reading metadata and data associated with that logical element, determining what data should be written durably into the storage system to persist any changes for the operation, or where metadata and data can be determined to be stored across modular storage devices attached to a plurality of the storage controllers in the storage system.
In some embodiments the operations are token based transactions to efficiently communicate within a distributed system. Each transaction may be accompanied by or associated with a token, which gives permission to execute the transaction. The authorities 168 are able to maintain a pre-transaction state of the system until completion of the operation in some embodiments. The token based communication may be accomplished without a global lock across the system, and also enables restart of an operation in case of a disruption or other failure.
In some systems, for example in UNIX-style file systems, data is handled with an index node or inode, which specifies a data structure that represents an object in a file system. The object could be a file or a directory, for example. Metadata may accompany the object, as attributes such as permission data and a creation timestamp, among other attributes. A segment number could be assigned to all or a portion of such an object in a file system. In other systems, data segments are handled with a segment number assigned elsewhere. For purposes of discussion, the unit of distribution is an entity, and an entity can be a file, a directory or a segment. That is, entities are units of data or metadata stored by a storage system. Entities are grouped into sets called authorities. Each authority has an authority owner, which is a storage node that has the exclusive right to update the entities in the authority. In other words, a storage node contains the authority, and that the authority, in turn, contains entities.
A segment is a logical container of data in accordance with some embodiments. A segment is an address space between medium address space and physical flash locations, i.e., the data segment number, are in this address space. Segments may also contain meta-data, which enable data redundancy to be restored (rewritten to different flash locations or devices) without the involvement of higher level software. In one embodiment, an internal format of a segment contains client data and medium mappings to determine the position of that data. Each data segment is protected, e.g., from memory and other failures, by breaking the segment into a number of data and parity shards, where applicable. The data and parity shards are distributed, i.e., striped, across non-volatile solid state storage 152 coupled to the host CPUs 156 (See
A series of address-space transformations takes place across an entire storage system. At the top are the directory entries (file names) which link to an inode. Inodes point into medium address space, where data is logically stored. Medium addresses may be mapped through a series of indirect mediums to spread the load of large files, or implement data services like deduplication or snapshots. Segment addresses are then translated into physical flash locations. Physical flash locations have an address range bounded by the amount of flash in the system in accordance with some embodiments. Medium addresses and segment addresses are logical containers, and in some embodiments use a 128 bit or larger identifier so as to be practically infinite, with a likelihood of reuse calculated as longer than the expected life of the system. Addresses from logical containers are allocated in a hierarchical fashion in some embodiments. Initially, each non-volatile solid state storage unit 152 may be assigned a range of address space. Within this assigned range, the non-volatile solid state storage 152 is able to allocate addresses without synchronization with other non-volatile solid state storage 152.
Data and metadata is stored by a set of underlying storage layouts that are optimized for varying workload patterns and storage devices. These layouts incorporate multiple redundancy schemes, compression formats and index algorithms. Some of these layouts store information about authorities and authority masters, while others store file metadata and file data. The redundancy schemes include error correction codes that tolerate corrupted bits within a single storage device (such as a NAND flash chip), erasure codes that tolerate the failure of multiple storage nodes, and replication schemes that tolerate data center or regional failures. In some embodiments, low density parity check (‘LDPC’) code is used within a single storage unit. Reed-Solomon encoding is used within a storage cluster, and mirroring is used within a storage grid in some embodiments. Metadata may be stored using an ordered log structured index (such as a Log Structured Merge Tree), and large data may not be stored in a log structured layout.
In order to maintain consistency across multiple copies of an entity, the storage nodes agree implicitly on two things through calculations: (1) the authority that contains the entity, and (2) the storage node that contains the authority. The assignment of entities to authorities can be done by pseudo randomly assigning entities to authorities, by splitting entities into ranges based upon an externally produced key, or by placing a single entity into each authority. Examples of pseudorandom schemes are linear hashing and the Replication Under Scalable Hashing (‘RUSH’) family of hashes, including Controlled Replication Under Scalable Hashing (‘CRUSH’). In some embodiments, pseudo-random assignment is utilized only for assigning authorities to nodes because the set of nodes can change. The set of authorities cannot change so any subjective function may be applied in these embodiments. Some placement schemes automatically place authorities on storage nodes, while other placement schemes rely on an explicit mapping of authorities to storage nodes. In some embodiments, a pseudorandom scheme is utilized to map from each authority to a set of candidate authority owners. A pseudorandom data distribution function related to CRUSH may assign authorities to storage nodes and create a list of where the authorities are assigned. Each storage node has a copy of the pseudorandom data distribution function, and can arrive at the same calculation for distributing, and later finding or locating an authority. Each of the pseudorandom schemes requires the reachable set of storage nodes as input in some embodiments in order to conclude the same target nodes. Once an entity has been placed in an authority, the entity may be stored on physical devices so that no expected failure will lead to unexpected data loss. In some embodiments, rebalancing algorithms attempt to store the copies of all entities within an authority in the same layout and on the same set of machines.
Examples of expected failures include device failures, stolen machines, datacenter fires, and regional disasters, such as nuclear or geological events. Different failures lead to different levels of acceptable data loss. In some embodiments, a stolen storage node impacts neither the security nor the reliability of the system, while depending on system configuration, a regional event could lead to no loss of data, a few seconds or minutes of lost updates, or even complete data loss.
In the embodiments, the placement of data for storage redundancy is independent of the placement of authorities for data consistency. In some embodiments, storage nodes that contain authorities do not contain any persistent storage. Instead, the storage nodes are connected to non-volatile solid state storage units that do not contain authorities. The communications interconnect between storage nodes and non-volatile solid state storage units consists of multiple communication technologies and has non-uniform performance and fault tolerance characteristics. In some embodiments, as mentioned above, non-volatile solid state storage units are connected to storage nodes via PCI express, storage nodes are connected together within a single chassis using Ethernet backplane, and chassis are connected together to form a storage cluster. Storage clusters are connected to clients using Ethernet or fiber channel in some embodiments. If multiple storage clusters are configured into a storage grid, the multiple storage clusters are connected using the Internet or other long-distance networking links, such as a “metro scale” link or private link that does not traverse the internet.
Authority owners have the exclusive right to modify entities, to migrate entities from one non-volatile solid state storage unit to another non-volatile solid state storage unit, and to add and remove copies of entities. This allows for maintaining the redundancy of the underlying data. When an authority owner fails, is going to be decommissioned, or is overloaded, the authority is transferred to a new storage node. Transient failures make it non-trivial to ensure that all non-faulty machines agree upon the new authority location. The ambiguity that arises due to transient failures can be achieved automatically by a consensus protocol such as Paxos, hot-warm failover schemes, via manual intervention by a remote system administrator, or by a local hardware administrator (such as by physically removing the failed machine from the cluster, or pressing a button on the failed machine). In some embodiments, a consensus protocol is used, and failover is automatic. If too many failures or replication events occur in too short a time period, the system goes into a self-preservation mode and halts replication and data movement activities until an administrator intervenes in accordance with some embodiments.
As authorities are transferred between storage nodes and authority owners update entities in their authorities, the system transfers messages between the storage nodes and non-volatile solid state storage units. With regard to persistent messages, messages that have different purposes are of different types. Depending on the type of the message, the system maintains different ordering and durability guarantees. As the persistent messages are being processed, the messages are temporarily stored in multiple durable and non-durable storage hardware technologies. In some embodiments, messages are stored in RAM, NVRAM and on NAND flash devices, and a variety of protocols are used in order to make efficient use of each storage medium. Latency-sensitive client requests may be persisted in replicated NVRAM, and then later NAND, while background rebalancing operations are persisted directly to NAND.
Persistent messages are persistently stored prior to being transmitted. This allows the system to continue to serve client requests despite failures and component replacement. Although many hardware components contain unique identifiers that are visible to system administrators, manufacturer, hardware supply chain and ongoing monitoring quality control infrastructure, applications running on top of the infrastructure address virtualize addresses. These virtualized addresses do not change over the lifetime of the storage system, regardless of component failures and replacements. This allows each component of the storage system to be replaced over time without reconfiguration or disruptions of client request processing, i.e., the system supports non-disruptive upgrades.
In some embodiments, the virtualized addresses are stored with sufficient redundancy. A continuous monitoring system correlates hardware and software status and the hardware identifiers. This allows detection and prediction of failures due to faulty components and manufacturing details. The monitoring system also enables the proactive transfer of authorities and entities away from impacted devices before failure occurs by removing the component from the critical path in some embodiments.
Storage clusters 161, in various embodiments as disclosed herein, can be contrasted with storage arrays in general. The storage nodes 150 are part of a collection that creates the storage cluster 161. Each storage node 150 owns a slice of data and computing required to provide the data. Multiple storage nodes 150 cooperate to store and retrieve the data. Storage memory or storage devices, as used in storage arrays in general, are less involved with processing and manipulating the data. Storage memory or storage devices in a storage array receive commands to read, write, or erase data. The storage memory or storage devices in a storage array are not aware of a larger system in which they are embedded, or what the data means. Storage memory or storage devices in storage arrays can include various types of storage memory, such as RAM, solid state drives, hard disk drives, etc. The storage units 152 described herein have multiple interfaces active simultaneously and serving multiple purposes. In some embodiments, some of the functionality of a storage node 150 is shifted into a storage unit 152, transforming the storage unit 152 into a combination of storage unit 152 and storage node 150. Placing computing (relative to storage data) into the storage unit 152 places this computing closer to the data itself. The various system embodiments have a hierarchy of storage node layers with different capabilities. By contrast, in a storage array, a controller owns and knows everything about all of the data that the controller manages in a shelf or storage devices. In a storage cluster 161, as described herein, multiple controllers in multiple storage units 152 and/or storage nodes 150 cooperate in various ways (e.g., for erasure coding, data sharding, metadata communication and redundancy, storage capacity expansion or contraction, data recovery, and so on).
The physical storage is divided into named regions based on application usage in some embodiments. The NVRAM 204 is a contiguous block of reserved memory in the storage unit 152 DRAM 216, and is backed by NAND flash. NVRAM 204 is logically divided into multiple memory regions written for two as spool (e.g., spool_region). Space within the NVRAM 204 spools is managed by each authority 168 independently. Each device provides an amount of storage space to each authority 168. That authority 168 further manages lifetimes and allocations within that space. Examples of a spool include distributed transactions or notions. When the primary power to a storage unit 152 fails, onboard super-capacitors provide a short duration of power hold up. During this holdup interval, the contents of the NVRAM 204 are flushed to flash memory 206. On the next power-on, the contents of the NVRAM 204 are recovered from the flash memory 206.
As for the storage unit controller, the responsibility of the logical “controller” is distributed across each of the blades containing authorities 168. This distribution of logical control is shown in
In the compute and storage planes 256, 258 of
Still referring to
Because authorities 168 are stateless, they can migrate between blades 252. Each authority 168 has a unique identifier. NVRAM 204 and flash 206 partitions are associated with authorities' 168 identifiers, not with the blades 252 on which they are running in some embodiments. Thus, when an authority 168 migrates, the authority 168 continues to manage the same storage partitions from its new location. When a new blade 252 is installed in an embodiment of the storage cluster, the system automatically rebalances load by: partitioning the new blade's 252 storage for use by the system's authorities 168, migrating selected authorities 168 to the new blade 252, starting endpoints 272 on the new blade 252 and including them in the switch fabric's 146 client connection distribution algorithm.
From their new locations, migrated authorities 168 persist the contents of their NVRAM 204 partitions on flash 206, process read and write requests from other authorities 168, and fulfill the client requests that endpoints 272 direct to them. Similarly, if a blade 252 fails or is removed, the system redistributes its authorities 168 among the system's remaining blades 252. The redistributed authorities 168 continue to perform their original functions from their new locations.
The embodiments described herein may utilize various software, communication and/or networking protocols. In addition, the configuration of the hardware and/or software may be adjusted to accommodate various protocols. For example, the embodiments may utilize Active Directory, which is a database based system that provides authentication, directory, policy, and other services in a WINDOWS™ environment. In these embodiments, LDAP (Lightweight Directory Access Protocol) is one example application protocol for querying and modifying items in directory service providers such as Active Directory. In some embodiments, a network lock manager (‘NLM’) is utilized as a facility that works in cooperation with the Network File System (‘NFS’) to provide a System V style of advisory file and record locking over a network. The Server Message Block (‘SMB’) protocol, one version of which is also known as Common
Internet File System (‘CIFS’), may be integrated with the storage systems discussed herein. SMP operates as an application-layer network protocol typically used for providing shared access to files, printers, and serial ports and miscellaneous communications between nodes on a network. SMB also provides an authenticated inter-process communication mechanism. AMAZON™ S3 (Simple Storage Service) is a web service offered by Amazon Web Services, and the systems described herein may interface with Amazon S3 through web services interfaces (REST (representational state transfer), SOAP (simple object access protocol), and BitTorrent). A RESTful API (application programming interface) breaks down a transaction to create a series of small modules. Each module addresses a particular underlying part of the transaction. The control or permissions provided with these embodiments, especially for object data, may include utilization of an access control list (‘ACL’). The ACL is a list of permissions attached to an object and the ACL specifies which users or system processes are granted access to objects, as well as what operations are allowed on given objects. The systems may utilize Internet Protocol version 6 (‘IPv6’), as well as IPv4, for the communications protocol that provides an identification and location system for computers on networks and routes traffic across the Internet. The routing of packets between networked systems may include Equal-cost multi-path routing (‘ECMP’), which is a routing strategy where next-hop packet forwarding to a single destination can occur over multiple “best paths” which tie for top place in routing metric calculations. Multi-path routing can be used in conjunction with most routing protocols, because it is a per-hop decision limited to a single router. The software may support Multi-tenancy, which is an architecture in which a single instance of a software application serves multiple customers. Each customer may be referred to as a tenant. Tenants may be given the ability to customize some parts of the application, but may not customize the application's code, in some embodiments. The embodiments may maintain audit logs. An audit log is a document that records an event in a computing system. In addition to documenting what resources were accessed, audit log entries typically include destination and source addresses, a timestamp, and user login information for compliance with various regulations. The embodiments may support various key management policies, such as encryption key rotation. In addition, the system may support dynamic root passwords or some variation dynamically changing passwords.
In the example depicted in
The cloud services provider 302 depicted in
In the example depicted in
In the example depicted in
Although not explicitly depicted in
In order to enable the storage system 306 and users of the storage system 306 to make use of the services provided by the cloud services provider 302, a cloud migration process may take place during which data, applications, or other elements from an organization's local systems (or even from another cloud environment) are moved to the cloud services provider 302. In order to successfully migrate data, applications, or other elements to the cloud services provider's 302 environment, middleware such as a cloud migration tool may be utilized to bridge gaps between the cloud services provider's 302 environment and an organization's environment. Such cloud migration tools may also be configured to address potentially high network costs and long transfer times associated with migrating large volumes of data to the cloud services provider 302, as well as addressing security concerns associated with sensitive data to the cloud services provider 302 over data communications networks. In order to further enable the storage system 306 and users of the storage system 306 to make use of the services provided by the cloud services provider 302, a cloud orchestrator may also be used to arrange and coordinate automated tasks in pursuit of creating a consolidated process or workflow. Such a cloud orchestrator may perform tasks such as configuring various components, whether those components are cloud components or on-premises components, as well as managing the interconnections between such components. The cloud orchestrator can simplify the inter-component communication and connections to ensure that links are correctly configured and maintained.
In the example depicted in
The cloud services provider 302 may also be configured to provide access to virtualized computing environments to the storage system 306 and users of the storage system 306. Such virtualized computing environments may be embodied, for example, as a virtual machine or other virtualized computer hardware platforms, virtual storage devices, virtualized computer network resources, and so on. Examples of such virtualized environments can include virtual machines that are created to emulate an actual computer, virtualized desktop environments that separate a logical desktop from a physical machine, virtualized file systems that allow uniform access to different types of concrete file systems, and many others.
Although the example depicted in
Readers will appreciate that by pairing the storage systems described herein with one or more cloud services providers, various offerings may be enabled. For example, disaster recovery as a service (‘DRaaS’) may be provided where cloud resources are utilized to protect applications and data from disruption caused by disaster, including in embodiments where the storage systems may serve as the primary data store. In such embodiments, a total system backup may be taken that allows for business continuity in the event of system failure. In such embodiments, cloud data backup techniques (by themselves or as part of a larger DRaaS solution) may also be integrated into an overall solution that includes the storage systems and cloud services providers described herein.
The storage systems described herein, as well as the cloud services providers, may be utilized to provide a wide array of security features. For example, the storage systems may encrypt data at rest (and data may be sent to and from the storage systems encrypted) and may make use of Key Management-as-a-Service (‘KMaaS’) to manage encryption keys, keys for locking and unlocking storage devices, and so on. Likewise, cloud data security gateways or similar mechanisms may be utilized to ensure that data stored within the storage systems does not improperly end up being stored in the cloud as part of a cloud data backup operation.
Furthermore, microsegmentation or identity-based-segmentation may be utilized in a data center that includes the storage systems or within the cloud services provider, to create secure zones in data centers and cloud deployments that enables the isolation of workloads from one another.
For further explanation,
The storage system 306 depicted in
The storage resources 308 depicted in
The storage resources 308 depicted in
The example storage system 306 depicted in
The example storage system 306 depicted in
The example storage system 306 depicted in
The storage system 306 depicted in
The communications resources 310 can also include mechanisms for accessing storage resources 308 within the storage system 306 utilizing serial attached SCSI (‘SAS’), serial ATA (‘SATA’) bus interfaces for connecting storage resources 308 within the storage system 306 to host bus adapters within the storage system 306, internet small computer systems interface (‘iSCSI’) technologies to provide block-level access to storage resources 308 within the storage system 306, and other communications resources that may be useful in facilitating data communications between components within the storage system 306, as well as data communications between the storage system 306 and computing devices that are outside of the storage system 306.
The storage system 306 depicted in
The storage system 306 depicted in
The software resources 314 may also include software that is useful in implementing software-defined storage (‘SDS’). In such an example, the software resources 314 may include one or more modules of computer program instructions that, when executed, are useful in policy-based provisioning and management of data storage that is independent of the underlying hardware. Such software resources 314 may be useful in implementing storage virtualization to separate the storage hardware from the software that manages the storage hardware.
The software resources 314 may also include software that is useful in facilitating and optimizing I/O operations that are directed to the storage resources 308 in the storage system 306. For example, the software resources 314 may include software modules that perform carry out various data reduction techniques such as, for example, data compression, data deduplication, and others. The software resources 314 may include software modules that intelligently group together I/O operations to facilitate better usage of the underlying storage resource 308, software modules that perform data migration operations to migrate from within a storage system, as well as software modules that perform other functions. Such software resources 314 may be embodied as one or more software containers or in many other ways.
For further explanation,
The cloud-based storage system 318 depicted in
In the example method depicted in
Consider an example in which the cloud computing environment 316 is embodied as AWS and the cloud computing instances are embodied as EC2 instances. In such an example, the cloud computing instance 320 that operates as the primary controller may be deployed on one of the instance types that has a relatively large amount of memory and processing power while the cloud computing instance 322 that operates as the secondary controller may be deployed on one of the instance types that has a relatively small amount of memory and processing power. In such an example, upon the occurrence of a failover event where the roles of primary and secondary are switched, a double failover may actually be carried out such that: 1) a first failover event where the cloud computing instance 322 that formerly operated as the secondary controller begins to operate as the primary controller, and 2) a third cloud computing instance (not shown) that is of an instance type that has a relatively large amount of memory and processing power is spun up with a copy of the storage controller application, where the third cloud computing instance begins operating as the primary controller while the cloud computing instance 322 that originally operated as the secondary controller begins operating as the secondary controller again. In such an example, the cloud computing instance 320 that formerly operated as the primary controller may be terminated. Readers will appreciate that in alternative embodiments, the cloud computing instance 320 that is operating as the secondary controller after the failover event may continue to operate as the secondary controller and the cloud computing instance 322 that operated as the primary controller after the occurrence of the failover event may be terminated once the primary role has been assumed by the third cloud computing instance (not shown).
Readers will appreciate that while the embodiments described above relate to embodiments where one cloud computing instance 320 operates as the primary controller and the second cloud computing instance 322 operates as the secondary controller, other embodiments are within the scope of the present disclosure. For example, each cloud computing instance 320, 322 may operate as a primary controller for some portion of the address space supported by the cloud-based storage system 318, each cloud computing instance 320, 322 may operate as a primary controller where the servicing of I/O operations directed to the cloud-based storage system 318 are divided in some other way, and so on. In fact, in other embodiments where costs savings may be prioritized over performance demands, only a single cloud computing instance may exist that contains the storage controller application.
The cloud-based storage system 318 depicted in
In the example depicted in
In the example depicted in
In the example depicted in
Readers will appreciate that when a request to write data is received by a particular cloud computing instance 340a, 340b, 340n with local storage 330, 334, 338, the software daemon 328, 332, 336 or some other module of computer program instructions that is executing on the particular cloud computing instance 340a, 340b, 340n may be configured to not only write the data to its own local storage 330, 334, 338 resources and any appropriate block-storage 342, 344, 346 that are offered by the cloud computing environment 316, but the software daemon 328, 332, 336 or some other module of computer program instructions that is executing on the particular cloud computing instance 340a, 340b, 340n may also be configured to write the data to cloud-based object storage 348 that is attached to the particular cloud computing instance 340a, 340b, 340n. The cloud-based object storage 348 that is attached to the particular cloud computing instance 340a, 340b, 340n may be embodied, for example, as Amazon Simple Storage Service (‘S3’) storage that is accessible by the particular cloud computing instance 340a, 340b, 340n. In other embodiments, the cloud computing instances 320, 322 that each include the storage controller application 324, 326 may initiate the storage of the data in the local storage 330, 334, 338 of the cloud computing instances 340a, 340b, 340n and the cloud-based object storage 348.
Readers will appreciate that, as described above, the cloud-based storage system 318 may be used to provide block storage services to users of the cloud-based storage system 318. While the local storage 330, 334, 338 resources and the block-storage 342, 344, 346 resources that are utilized by the cloud computing instances 340a, 340b, 340n may support block-level access, the cloud-based object storage 348 that is attached to the particular cloud computing instance 340a, 340b, 340n supports only object-based access. In order to address this, the software daemon 328, 332, 336 or some other module of computer program instructions that is executing on the particular cloud computing instance 340a, 340b, 340n may be configured to take blocks of data, package those blocks into objects, and write the objects to the cloud-based object storage 348 that is attached to the particular cloud computing instance 340a, 340b, 340n.
Consider an example in which data is written to the local storage 330, 334, 338 resources and the block-storage 342, 344, 346 resources that are utilized by the cloud computing instances 340a, 340b, 340n in 1 MB blocks. In such an example, assume that a user of the cloud-based storage system 318 issues a request to write data that, after being compressed and deduplicated by the storage controller application 324, 326 results in the need to write 5 MB of data. In such an example, writing the data to the local storage 330, 334, 338 resources and the block-storage 342, 344, 346 resources that are utilized by the cloud computing instances 340a, 340b, 340n is relatively straightforward as 5 blocks that are 1 MB in size are written to the local storage 330, 334, 338 resources and the block-storage 342, 344, 346 resources that are utilized by the cloud computing instances 340a, 340b, 340n. In such an example, the software daemon 328, 332, 336 or some other module of computer program instructions that is executing on the particular cloud computing instance 340a, 340b, 340n may be configured to: 1) create a first object that includes the first 1 MB of data and write the first object to the cloud-based object storage 348, 2) create a second object that includes the second 1 MB of data and write the second object to the cloud-based object storage 348, 3) create a third object that includes the third 1 MB of data and write the third object to the cloud-based object storage 348, and so on. As such, in some embodiments, each object that is written to the cloud-based object storage 348 may be identical (or nearly identical) in size. Readers will appreciate that in such an example, metadata that is associated with the data itself may be included in each object (e.g., the first 1 MB of the object is data and the remaining portion is metadata associated with the data).
Readers will appreciate that the cloud-based object storage 348 may be incorporated into the cloud-based storage system 318 to increase the durability of the cloud-based storage system 318. Continuing with the example described above where the cloud computing instances 340a, 340b, 340n are EC2 instances, readers will understand that EC2 instances are only guaranteed to have a monthly uptime of 99.9% and data stored in the local instance store only persists during the lifetime of the EC2 instance. As such, relying on the cloud computing instances 340a, 340b, 340n with local storage 330, 334, 338 as the only source of persistent data storage in the cloud-based storage system 318 may result in a relatively unreliable storage system. Likewise, EBS volumes are designed for 99.999% availability. As such, even relying on EBS as the persistent data store in the cloud-based storage system 318 may result in a storage system that is not sufficiently durable. Amazon S3, however, is designed to provide 99.999999999% durability, meaning that a cloud-based storage system 318 that can incorporate S3 into its pool of storage is substantially more durable than various other options.
Readers will appreciate that while a cloud-based storage system 318 that can incorporate S3 into its pool of storage is substantially more durable than various other options, utilizing S3 as the primary pool of storage may result in storage system that has relatively slow response times and relatively long I/O latencies. As such, the cloud-based storage system 318 depicted in
In some embodiments, all data that is stored by the cloud-based storage system 318 may be stored in both: 1) the cloud-based object storage 348, and 2) at least one of the local storage 330, 334, 338 resources or block-storage 342, 344, 346 resources that are utilized by the cloud computing instances 340a, 340b, 340n. In such embodiments, the local storage 330, 334, 338 resources and block-storage 342, 344, 346 resources that are utilized by the cloud computing instances 340a, 340b, 340n may effectively operate as cache that generally includes all data that is also stored in S3, such that all reads of data may be serviced by the cloud computing instances 340a, 340b, 340n without requiring the cloud computing instances 340a, 340b, 340n to access the cloud-based object storage 348. Readers will appreciate that in other embodiments, however, all data that is stored by the cloud-based storage system 318 may be stored in the cloud-based object storage 348, but less than all data that is stored by the cloud-based storage system 318 may be stored in at least one of the local storage 330, 334, 338 resources or block-storage 342, 344, 346 resources that are utilized by the cloud computing instances 340a, 340b, 340n. In such an example, various policies may be utilized to determine which subset of the data that is stored by the cloud-based storage system 318 should reside in both: 1) the cloud-based object storage 348, and 2) at least one of the local storage 330, 334, 338 resources or block-storage 342, 344, 346 resources that are utilized by the cloud computing instances 340a, 340b, 340n.
As described above, when the cloud computing instances 340a, 340b, 340n with local storage 330, 334, 338 are embodied as EC2 instances, the cloud computing instances 340a, 340b, 340n with local storage 330, 334, 338 are only guaranteed to have a monthly uptime of 99.9% and data stored in the local instance store only persists during the lifetime of each cloud computing instance 340a, 340b, 340n with local storage 330, 334, 338. As such, one or more modules of computer program instructions that are executing within the cloud-based storage system 318 (e.g., a monitoring module that is executing on its own EC2 instance) may be designed to handle the failure of one or more of the cloud computing instances 340a, 340b, 340n with local storage 330, 334, 338. In such an example, the monitoring module may handle the failure of one or more of the cloud computing instances 340a, 340b, 340n with local storage 330, 334, 338 by creating one or more new cloud computing instances with local storage, retrieving data that was stored on the failed cloud computing instances 340a, 340b, 340n from the cloud-based object storage 348, and storing the data retrieved from the cloud-based object storage 348 in local storage on the newly created cloud computing instances. Readers will appreciate that many variants of this process may be implemented.
Consider an example in which all cloud computing instances 340a, 340b, 340n with local storage 330, 334, 338 failed. In such an example, the monitoring module may create new cloud computing instances with local storage, where high-bandwidth instances types are selected that allow for the maximum data transfer rates between the newly created high-bandwidth cloud computing instances with local storage and the cloud-based object storage 348. Readers will appreciate that instances types are selected that allow for the maximum data transfer rates between the new cloud computing instances and the cloud-based object storage 348 such that the new high-bandwidth cloud computing instances can be rehydrated with data from the cloud-based object storage 348 as quickly as possible. Once the new high-bandwidth cloud computing instances are rehydrated with data from the cloud-based object storage 348, less expensive lower-bandwidth cloud computing instances may be created, data may be migrated to the less expensive lower-bandwidth cloud computing instances, and the high-bandwidth cloud computing instances may be terminated.
Readers will appreciate that in some embodiments, the number of new cloud computing instances that are created may substantially exceed the number of cloud computing instances that are needed to locally store all of the data stored by the cloud-based storage system 318. The number of new cloud computing instances that are created may substantially exceed the number of cloud computing instances that are needed to locally store all of the data stored by the cloud-based storage system 318 in order to more rapidly pull data from the cloud-based object storage 348 and into the new cloud computing instances, as each new cloud computing instance can (in parallel) retrieve some portion of the data stored by the cloud-based storage system 318. In such embodiments, once the data stored by the cloud-based storage system 318 has been pulled into the newly created cloud computing instances, the data may be consolidated within a subset of the newly created cloud computing instances and those newly created cloud computing instances that are excessive may be terminated.
Consider an example in which 1000 cloud computing instances are needed in order to locally store all valid data that users of the cloud-based storage system 318 have written to the cloud-based storage system 318. In such an example, assume that all 1,000 cloud computing instances fail. In such an example, the monitoring module may cause 100,000 cloud computing instances to be created, where each cloud computing instance is responsible for retrieving, from the cloud-based object storage 348, distinct 1/100,000th chunks of the valid data that users of the cloud-based storage system 318 have written to the cloud-based storage system 318 and locally storing the distinct chunk of the dataset that it retrieved. In such an example, because each of the 100,000 cloud computing instances can retrieve data from the cloud-based object storage 348 in parallel, the caching layer may be restored 100 times faster as compared to an embodiment where the monitoring module only create 1000 replacement cloud computing instances. In such an example, over time the data that is stored locally in the 100,000 could be consolidated into 1,000cloud computing instances and the remaining 99,000 cloud computing instances could be terminated.
Readers will appreciate that various performance aspects of the cloud-based storage system 318 may be monitored (e.g., by a monitoring module that is executing in an EC2 instance) such that the cloud-based storage system 318 can be scaled-up or scaled-out as needed. Consider an example in which the monitoring module monitors the performance of the could-based storage system 318 via communications with one or more of the cloud computing instances 320, 322 that each are used to support the execution of a storage controller application 324, 326, via monitoring communications between cloud computing instances 320, 322, 340a, 340b, 340n, via monitoring communications between cloud computing instances 320, 322, 340a, 340b, 340n and the cloud-based object storage 348, or in some other way. In such an example, assume that the monitoring module determines that the cloud computing instances 320, 322 that are used to support the execution of a storage controller application 324, 326 are undersized and not sufficiently servicing the I/O requests that are issued by users of the cloud-based storage system 318. In such an example, the monitoring module may create a new, more powerful cloud computing instance (e.g., a cloud computing instance of a type that includes more processing power, more memory, etc. . . . ) that includes the storage controller application such that the new, more powerful cloud computing instance can begin operating as the primary controller.
Likewise, if the monitoring module determines that the cloud computing instances 320, 322 that are used to support the execution of a storage controller application 324, 326 are oversized and that cost savings could be gained by switching to a smaller, less powerful cloud computing instance, the monitoring module may create a new, less powerful (and less expensive) cloud computing instance that includes the storage controller application such that the new, less powerful cloud computing instance can begin operating as the primary controller.
Consider, as an additional example of dynamically sizing the cloud-based storage system 318, an example in which the monitoring module determines that the utilization of the local storage that is collectively provided by the cloud computing instances 340a, 340b, 340n has reached a predetermined utilization threshold (e.g., 95%). In such an example, the monitoring module may create additional cloud computing instances with local storage to expand the pool of local storage that is offered by the cloud computing instances. Alternatively, the monitoring module may create one or more new cloud computing instances that have larger amounts of local storage than the already existing cloud computing instances 340a, 340b, 340n, such that data stored in an already existing cloud computing instance 340a, 340b, 340n can be migrated to the one or more new cloud computing instances and the already existing cloud computing instance 340a, 340b, 340n can be terminated, thereby expanding the pool of local storage that is offered by the cloud computing instances. Likewise, if the pool of local storage that is offered by the cloud computing instances is unnecessarily large, data can be consolidated and some cloud computing instances can be terminated.
Readers will appreciate that the cloud-based storage system 318 may be sized up and down automatically by a monitoring module applying a predetermined set of rules that may be relatively simple of relatively complicated. In fact, the monitoring module may not only take into account the current state of the cloud-based storage system 318, but the monitoring module may also apply predictive policies that are based on, for example, observed behavior (e.g., every night from 10 PM until 6 AM usage of the storage system is relatively light), predetermined fingerprints (e.g., every time a virtual desktop infrastructure adds 100 virtual desktops, the number of IOPS directed to the storage system increase by X), and so on. In such an example, the dynamic scaling of the cloud-based storage system 318 may be based on current performance metrics, predicted workloads, and many other factors, including combinations thereof.
Readers will further appreciate that because the cloud-based storage system 318 may be dynamically scaled, the cloud-based storage system 318 may even operate in a way that is more dynamic. Consider the example of garbage collection. In a traditional storage system, the amount of storage is fixed. As such, at some point the storage system may be forced to perform garbage collection as the amount of available storage has become so constrained that the storage system is on the verge of running out of storage. In contrast, the cloud-based storage system 318 described here can always ‘add’ additional storage (e.g., by adding more cloud computing instances with local storage). Because the cloud-based storage system 318 described here can always ‘add’ additional storage, the cloud-based storage system 318 can make more intelligent decisions regarding when to perform garbage collection. For example, the cloud-based storage system 318 may implement a policy that garbage collection only be performed when the number of IOPS being serviced by the cloud-based storage system 318 falls below a certain level. In some embodiments, other system-level functions (e.g., deduplication, compression) may also be turned off and on in response to system load, given that the size of the cloud-based storage system 318 is not constrained in the same way that traditional storage systems are constrained.
Readers will appreciate that embodiments of the present disclosure resolve an issue with block-storage services offered by some cloud computing environments as some cloud computing environments only allow for one cloud computing instance to connect to a block-storage volume at a single time. For example, in Amazon AWS, only a single EC2 instance may be connected to an EBS volume. Through the use of EC2 instances with local storage, embodiments of the present disclosure can offer multi-connect capabilities where multiple EC2 instances can connect to another EC2 instance with local storage (‘a drive instance’). In such embodiments, the drive instances may include software executing within the drive instance that allows the drive instance to support I/O directed to a particular volume from each connected EC2 instance. As such, some embodiments of the present disclosure may be embodied as multi-connect block storage services that may not include all of the components depicted in
In some embodiments, especially in embodiments where the cloud-based object storage 348 resources are embodied as Amazon S3, the cloud-based storage system 318 may include one or more modules (e.g., a module of computer program instructions executing on an EC2instance) that are configured to ensure that when the local storage of a particular cloud computing instance is rehydrated with data from S3, the appropriate data is actually in S3. This issue arises largely because S3 implements an eventual consistency model where, when overwriting an existing object, reads of the object will eventually (but not necessarily immediately) become consistent and will eventually (but not necessarily immediately) return the overwritten version of the object. To address this issue, in some embodiments of the present disclosure, objects in S3 are never overwritten. Instead, a traditional ‘overwrite’ would result in the creation of the new object (that includes the updated version of the data) and the eventual deletion of the old object (that includes the previous version of the data).
In some embodiments of the present disclosure, as part of an attempt to never (or almost never) overwrite an object, when data is written to S3 the resultant object may be tagged with a sequence number. In some embodiments, these sequence numbers may be persisted elsewhere (e.g., in a database) such that at any point in time, the sequence number associated with the most up-to-date version of some piece of data can be known. In such a way, a determination can be made as to whether S3 has the most recent version of some piece of data by merely reading the sequence number associated with an object-and without actually reading the data from S3. The ability to make this determination may be particularly important when a cloud computing instance with local storage crashes, as it would be undesirable to rehydrate the local storage of a replacement cloud computing instance with out-of-date data. In fact, because the cloud-based storage system 318 does not need to access the data to verify its validity, the data can stay encrypted and access charges can be avoided.
The storage systems described above may carry out intelligent data backup techniques through which data stored in the storage system may be copied and stored in a distinct location to avoid data loss in the event of equipment failure or some other form of catastrophe. For example, the storage systems described above may be configured to examine each backup to avoid restoring the storage system to an undesirable state. Consider an example in which malware infects the storage system. In such an example, the storage system may include software resources 314 that can scan each backup to identify backups that were captured before the malware infected the storage system and those backups that were captured after the malware infected the storage system. In such an example, the storage system may restore itself from a backup that does not include the malware-or at least not restore the portions of a backup that contained the malware. In such an example, the storage system may include software resources 314 that can scan each backup to identify the presences of malware (or a virus, or some other undesirable), for example, by identifying write operations that were serviced by the storage system and originated from a network subnet that is suspected to have delivered the malware, by identifying write operations that were serviced by the storage system and originated from a user that is suspected to have delivered the malware, by identifying write operations that were serviced by the storage system and examining the content of the write operation against fingerprints of the malware, and in many other ways.
Readers will further appreciate that the backups (often in the form of one or more snapshots) may also be utilized to perform rapid recovery of the storage system. Consider an example in which the storage system is infected with ransomware that locks users out of the storage system. In such an example, software resources 314 within the storage system may be configured to detect the presence of ransomware and may be further configured to restore the storage system to a point-in-time, using the retained backups, prior to the point-in-time at which the ransomware infected the storage system. In such an example, the presence of ransomware may be explicitly detected through the use of software tools utilized by the system, through the use of a key (e.g., a USB drive) that is inserted into the storage system, or in a similar way. Likewise, the presence of ransomware may be inferred in response to system activity meeting a predetermined fingerprint such as, for example, no reads or writes coming into the system for a predetermined period of time.
Readers will appreciate that the various components described above may be grouped into one or more optimized computing packages as converged infrastructures. Such converged infrastructures may include pools of computers, storage and networking resources that can be shared by multiple applications and managed in a collective manner using policy-driven processes. Such converged infrastructures may be implemented with a converged infrastructure reference architecture, with standalone appliances, with a software driven hyper-converged approach (e.g., hyper-converged infrastructures), or in other ways.
Readers will appreciate that the storage systems described above may be useful for supporting various types of software applications. For example, the storage system 306 may be useful in supporting artificial intelligence (‘AI’) applications, database applications, DevOps projects, electronic design automation tools, event-driven software applications, high performance computing applications, simulation applications, high-speed data capture and analysis applications, machine learning applications, media production applications, media serving applications, picture archiving and communication systems (‘PACS’) applications, software development applications, virtual reality applications, augmented reality applications, and many other types of applications by providing storage resources to such applications.
The storage systems described above may operate to support a wide variety of applications. In view of the fact that the storage systems include compute resources, storage resources, and a wide variety of other resources, the storage systems may be well suited to support applications that are resource intensive such as, for example, AI applications. AI applications may be deployed in a variety of fields, including: predictive maintenance in manufacturing and related fields, healthcare applications such as patient data & risk analytics, retail and marketing deployments (e.g., search advertising, social media advertising), supply chains solutions, fintech solutions such as business analytics & reporting tools, operational deployments such as real-time analytics tools, application performance management tools, IT infrastructure management tools, and many others.
Such AI applications may enable devices to perceive their environment and take actions that maximize their chance of success at some goal. Examples of such AI applications can include IBM Watson, Microsoft Oxford, Google DeepMind, Baidu Minwa, and others. The storage systems described above may also be well suited to support other types of applications that are resource intensive such as, for example, machine learning applications. Machine learning applications may perform various types of data analysis to automate analytical model building. Using algorithms that iteratively learn from data, machine learning applications can enable computers to learn without being explicitly programmed. One particular area of machine learning is referred to as reinforcement learning, which involves taking suitable actions to maximize reward in a particular situation. Reinforcement learning may be employed to find the best possible behavior or path that a particular software application or machine should take in a specific situation. Reinforcement learning differs from other areas of machine learning (e.g., supervised learning, unsupervised learning) in that correct input/output pairs need not be presented for reinforcement learning and sub-optimal actions need not be explicitly corrected.
In addition to the resources already described, the storage systems described above may also include graphics processing units (‘GPUs’), occasionally referred to as visual processing unit (‘VPUs’). Such GPUs may be embodied as specialized electronic circuits that rapidly manipulate and alter memory to accelerate the creation of images in a frame buffer intended for output to a display device. Such GPUs may be included within any of the computing devices that are part of the storage systems described above, including as one of many individually scalable components of a storage system, where other examples of individually scalable components of such storage system can include storage components, memory components, compute components (e.g., CPUs, FPGAs, ASICs), networking components, software components, and others. In addition to GPUs, the storage systems described above may also include neural network processors (‘NNPs’) for use in various aspects of neural network processing. Such NNPs may be used in place of (or in addition to) GPUs and may be also be independently scalable.
As described above, the storage systems described herein may be configured to support artificial intelligence applications, machine learning applications, big data analytics applications, and many other types of applications. The rapid growth in these sort of applications is being driven by three technologies: deep learning (DL), GPU processors, and Big Data. Deep learning is a computing model that makes use of massively parallel neural networks inspired by the human brain. Instead of experts handcrafting software, a deep learning model writes its own software by learning from lots of examples. Such GPUs may include thousands of cores that are well-suited to run algorithms that loosely represent the parallel nature of the human brain.
Advances in deep neural networks, including the development of multi-layer neural networks, have ignited a new wave of algorithms and tools for data scientists to tap into their data with artificial intelligence (AI). With improved algorithms, larger data sets, and various frameworks (including open-source software libraries for machine learning across a range of tasks), data scientists are tackling new use cases like autonomous driving vehicles, natural language processing and understanding, computer vision, machine reasoning, strong AI, and many others. Applications of such techniques may include: machine and vehicular object detection, identification and avoidance; visual recognition, classification and tagging; algorithmic financial trading strategy performance management; simultaneous localization and mapping; predictive maintenance of high-value machinery; prevention against cyber security threats, expertise automation; image recognition and classification; question answering; robotics; text analytics (extraction, classification) and text generation and translation; and many others. Applications of AI techniques has materialized in a wide array of products include, for example, Amazon Echo's speech recognition technology that allows users to talk to their machines, Google Translate™ which allows for machine-based language translation, Spotify's Discover Weekly that provides recommendations on new songs and artists that a user may like based on the user's usage and traffic analysis, Quill's text generation offering that takes structured data and turns it into narrative stories, Chatbots that provide real-time, contextually specific answers to questions in a dialog format, and many others.
Data is the heart of modern AI and deep learning algorithms. Before training can begin, one problem that must be addressed revolves around collecting the labeled data that is crucial for training an accurate AI model. A full scale AI deployment may be required to continuously collect, clean, transform, label, and store large amounts of data. Adding additional high quality data points directly translates to more accurate models and better insights. Data samples may undergo a series of processing steps including, but not limited to: 1) ingesting the data from an external source into the training system and storing the data in raw form, 2) cleaning and transforming the data in a format convenient for training, including linking data samples to the appropriate label, 3) exploring parameters and models, quickly testing with a smaller dataset, and iterating to converge on the most promising models to push into the production cluster, 4) executing training phases to select random batches of input data, including both new and older samples, and feeding those into production GPU servers for computation to update model parameters, and 5) evaluating including using a holdback portion of the data not used in training in order to evaluate model accuracy on the holdout data. This lifecycle may apply for any type of parallelized machine learning, not just neural networks or deep learning. For example, standard machine learning frameworks may rely on CPUs instead of GPUs but the data ingest and training workflows may be the same. Readers will appreciate that a single shared storage data hub creates a coordination point throughout the lifecycle without the need for extra data copies among the ingest, preprocessing, and training stages. Rarely is the ingested data used for only one purpose, and shared storage gives the flexibility to train multiple different models or apply traditional analytics to the data.
Readers will appreciate that each stage in the AI data pipeline may have varying requirements from the data hub (e.g., the storage system or collection of storage systems). Scale-out storage systems must deliver uncompromising performance for all manner of access types and patterns-from small, metadata-heavy to large files, from random to sequential access patterns, and from low to high concurrency. The storage systems described above may serve as an ideal AI data hub as the systems may service unstructured workloads. In the first stage, data is ideally ingested and stored on to the same data hub that following stages will use, in order to avoid excess data copying. The next two steps can be done on a standard compute server that optionally includes a GPU, and then in the fourth and last stage, full training production jobs are run on powerful GPU-accelerated servers. Often, there is a production pipeline alongside an experimental pipeline operating on the same dataset. Further, the GPU-accelerated servers can be used independently for different models or joined together to train on one larger model, even spanning multiple systems for distributed training. If the shared storage tier is slow, then data must be copied to local storage for each phase, resulting in wasted time staging data onto different servers. The ideal data hub for the AI training pipeline delivers performance similar to data stored locally on the server node while also having the simplicity and performance to enable all pipeline stages to operate concurrently.
In order for the storage systems described above to serve as a data hub or as part of an AI deployment, in some embodiments the storage systems may be configured to provide DMA between storage devices that are included in the storage systems and one or more GPUs that are used in an AI or big data analytics pipeline. The one or more GPUs may be coupled to the storage system, for example, via NVMe-over-Fabrics (‘NVMe-oF’) such that bottlenecks such as the host CPU can be bypassed and the storage system (or one of the components contained therein) can directly access GPU memory. In such an example, the storage systems may leverage API hooks to the GPUs to transfer data directly to the GPUs. For example, the GPUs may be embodied as Nvidia™ GPUs and the storage systems may support GPUDirect Storage (‘GDS’) software, or have similar proprietary software, that enables the storage system to transfer data to the GPUs via RDMA or similar mechanism. Readers will appreciate that in embodiments where the storage systems are embodied as cloud-based storage systems as described below, virtual drive or other components within such a cloud-based storage system may also be configured
Although the preceding paragraphs discuss deep learning applications, readers will appreciate that the storage systems described herein may also be part of a distributed deep learning (‘DDL’) platform to support the execution of DDL algorithms. The storage systems described above may also be paired with other technologies such as TensorFlow, an open-source software library for dataflow programming across a range of tasks that may be used for machine learning applications such as neural networks, to facilitate the development of such machine learning models, applications, and so on.
The storage systems described above may also be used in a neuromorphic computing environment. Neuromorphic computing is a form of computing that mimics brain cells. To support neuromorphic computing, an architecture of interconnected “neurons” replace traditional computing models with low-powered signals that go directly between neurons for more efficient computation. Neuromorphic computing may make use of very-large-scale integration (VLSI) systems containing electronic analog circuits to mimic neuro-biological architectures present in the nervous system, as well as analog, digital, mixed-mode analog/digital VLSI, and software systems that implement models of neural systems for perception, motor control, or multisensory integration.
Readers will appreciate that the storage systems described above may be configured to support the storage or use of (among other types of data) blockchains. In addition to supporting the storage and use of blockchain technologies, the storage systems described above may also support the storage and use of derivative items such as, for example, open source blockchains and related tools that are part of the IBM™ Hyperledger project, permissioned blockchains in which a certain number of trusted parties are allowed to access the block chain, blockchain products that enable developers to build their own distributed ledger projects, and others. Blockchains and the storage systems described herein may be leveraged to support on-chain storage of data as well as off-chain storage of data.
Off-chain storage of data can be implemented in a variety of ways and can occur when the data itself is not stored within the blockchain. For example, in one embodiment, a hash function may be utilized and the data itself may be fed into the hash function to generate a hash value. In such an example, the hashes of large pieces of data may be embedded within transactions, instead of the data itself. Readers will appreciate that, in other embodiments, alternatives to blockchains may be used to facilitate the decentralized storage of information. For example, one alternative to a blockchain that may be used is a blockweave. While conventional blockchains store every transaction to achieve validation, a blockweave permits secure decentralization without the usage of the entire chain, thereby enabling low cost on-chain storage of data. Such blockweaves may utilize a consensus mechanism that is based on proof of access (PoA) and proof of work (PoW).
The storage systems described above may, either alone or in combination with other computing devices, be used to support in-memory computing applications. In-memory computing involves the storage of information in RAM that is distributed across a cluster of computers. Readers will appreciate that the storage systems described above, especially those that are configurable with customizable amounts of processing resources, storage resources, and memory resources (e.g., those systems in which blades that contain configurable amounts of each type of resource), may be configured in a way so as to provide an infrastructure that can support in-memory computing. Likewise, the storage systems described above may include component parts (e.g., NVDIMMs, 3D crosspoint storage that provide fast random access memory that is persistent) that can actually provide for an improved in-memory computing environment as compared to in-memory computing environments that rely on RAM distributed across dedicated servers.
In some embodiments, the storage systems described above may be configured to operate as a hybrid in-memory computing environment that includes a universal interface to all storage media (e.g., RAM, flash storage, 3D crosspoint storage). In such embodiments, users may have no knowledge regarding the details of where their data is stored but they can still use the same full, unified API to address data. In such embodiments, the storage system may (in the background) move data to the fastest layer available-including intelligently placing the data in dependence upon various characteristics of the data or in dependence upon some other heuristic. In such an example, the storage systems may even make use of existing products such as Apache Ignite and GridGain to move data between the various storage layers, or the storage systems may make use of custom software to move data between the various storage layers. The storage systems described herein may implement various optimizations to improve the performance of in-memory computing such as, for example, having computations occur as close to the data as possible.
Readers will further appreciate that in some embodiments, the storage systems described above may be paired with other resources to support the applications described above. For example, one infrastructure could include primary compute in the form of servers and workstations which specialize in using General-purpose computing on graphics processing units (‘GPGPU’) to accelerate deep learning applications that are interconnected into a computation engine to train parameters for deep neural networks. Each system may have Ethernet external connectivity, InfiniBand external connectivity, some other form of external connectivity, or some combination thereof. In such an example, the GPUs can be grouped for a single large training or used independently to train multiple models. The infrastructure could also include a storage system such as those described above to provide, for example, a scale-out all-flash file or object store through which data can be accessed via high-performance protocols such as NFS, S3, and so on. The infrastructure can also include, for example, redundant top-of-rack Ethernet switches connected to storage and compute via ports in MLAG port channels for redundancy. The infrastructure could also include additional compute in the form of whitebox servers, optionally with GPUs, for data ingestion, pre-processing, and model debugging. Readers will appreciate that additional infrastructures are also possible.
Readers will appreciate that the storage systems described above, either alone or in coordination with other computing machinery may be configured to support other AI related tools. For example, the storage systems may make use of tools like ONXX or other open neural network exchange formats that make it easier to transfer models written in different AI frameworks. Likewise, the storage systems may be configured to support tools like Amazon's Gluon that allow developers to prototype, build, and train deep learning models. In fact, the storage systems described above may be part of a larger platform, such as IBM™ Cloud Private for Data, that includes integrated data science, data engineering and application building services.
Readers will further appreciate that the storage systems described above may also be deployed as an edge solution. Such an edge solution may be in place to optimize cloud computing systems by performing data processing at the edge of the network, near the source of the data. Edge computing can push applications, data and computing power (i.e., services) away from centralized points to the logical extremes of a network. Through the use of edge solutions such as the storage systems described above, computational tasks may be performed using the compute resources provided by such storage systems, data may be storage using the storage resources of the storage system, and cloud-based services may be accessed through the use of various resources of the storage system (including networking resources). By performing computational tasks on the edge solution, storing data on the edge solution, and generally making use of the edge solution, the consumption of expensive cloud-based resources may be avoided and, in fact, performance improvements may be experienced relative to a heavier reliance on cloud-based resources.
While many tasks may benefit from the utilization of an edge solution, some particular uses may be especially suited for deployment in such an environment. For example, devices like drones, autonomous cars, robots, and others may require extremely rapid processing-so fast, in fact, that sending data up to a cloud environment and back to receive data processing support may simply be too slow. As an additional example, some IoT devices such as connected video cameras may not be well-suited for the utilization of cloud-based resources as it may be impractical (not only from a privacy perspective, security perspective, or a financial perspective) to send the data to the cloud simply because of the pure volume of data that is involved. As such, many tasks that really on data processing, storage, or communications may be better suited by platforms that include edge solutions such as the storage systems described above.
The storage systems described above may alone, or in combination with other computing resources, serves as a network edge platform that combines compute resources, storage resources, networking resources, cloud technologies and network virtualization technologies, and so on. As part of the network, the edge may take on characteristics similar to other network facilities, from the customer premise and backhaul aggregation facilities to Points of Presence (PoPs) and regional data centers. Readers will appreciate that network workloads, such as Virtual Network Functions (VNFs) and others, will reside on the network edge platform. Enabled by a combination of containers and virtual machines, the network edge platform may rely on controllers and schedulers that are no longer geographically co-located with the data processing resources. The functions, as microservices, may split into control planes, user and data planes, or even state machines, allowing for independent optimization and scaling techniques to be applied. Such user and data planes may be enabled through increased accelerators, both those residing in server platforms, such as FPGAs and Smart NICs, and through SDN-enabled merchant silicon and programmable ASICs.
The storage systems described above may also be optimized for use in big data analytics. Big data analytics may be generally described as the process of examining large and varied data sets to uncover hidden patterns, unknown correlations, market trends, customer preferences and other useful information that can help organizations make more-informed business decisions. As part of that process, semi-structured and unstructured data such as, for example, internet clickstream data, web server logs, social media content, text from customer emails and survey responses, mobile-phone call-detail records, IoT sensor data, and other data may be converted to a structured form.
The storage systems described above may also support (including implementing as a system interface) applications that perform tasks in response to human speech. For example, the storage systems may support the execution intelligent personal assistant applications such as, for example, Amazon's Alexa, Apple Siri, Google Voice, Samsung Bixby, Microsoft Cortana, and others. While the examples described in the previous sentence make use of voice as input, the storage systems described above may also support chatbots, talkbots, chatterbots, or artificial conversational entities or other applications that are configured to conduct a conversation via auditory or textual methods. Likewise, the storage system may actually execute such an application to enable a user such as a system administrator to interact with the storage system via speech. Such applications are generally capable of voice interaction, music playback, making to-do lists, setting alarms, streaming podcasts, playing audiobooks, and providing weather, traffic, and other real time information, such as news, although in embodiments in accordance with the present disclosure, such applications may be utilized as interfaces to various system management operations.
The storage systems described above may also implement AI platforms for delivering on the vision of self-driving storage. Such AI platforms may be configured to deliver global predictive intelligence by collecting and analyzing large amounts of storage system telemetry data points to enable effortless management, analytics and support. In fact, such storage systems may be capable of predicting both capacity and performance, as well as generating intelligent advice on workload deployment, interaction and optimization. Such AI platforms may be configured to scan all incoming storage system telemetry data against a library of issue fingerprints to predict and resolve incidents in real-time, before they impact customer environments, and captures hundreds of variables related to performance that are used to forecast performance load.
The storage systems described above may support the serialized or simultaneous execution of artificial intelligence applications, machine learning applications, data analytics applications, data transformations, and other tasks that collectively may form an AI ladder. Such an AI ladder may effectively be formed by combining such elements to form a complete data science pipeline, where exist dependencies between elements of the AI ladder. For example, AI may require that some form of machine learning has taken place, machine learning may require that some form of analytics has taken place, analytics may require that some form of data and information architecting has taken place, and so on. As such, each element may be viewed as a rung in an AI ladder that collectively can form a complete and sophisticated AI solution.
The storage systems described above may also, either alone or in combination with other computing environments, be used to deliver an AI everywhere experience where AI permeates wide and expansive aspects of business and life. For example, AI may play an important role in the delivery of deep learning solutions, deep reinforcement learning solutions, artificial general intelligence solutions, autonomous vehicles, cognitive computing solutions, commercial UAVs or drones, conversational user interfaces, enterprise taxonomies, ontology management solutions, machine learning solutions, smart dust, smart robots, smart workplaces, and many others.
The storage systems described above may also, either alone or in combination with other computing environments, be used to deliver a wide range of transparently immersive experiences (including those that use digital twins of various “things” such as people, places, processes, systems, and so on) where technology can introduce transparency between people, businesses, and things. Such transparently immersive experiences may be delivered as augmented reality technologies, connected homes, virtual reality technologies, brain-computer interfaces, human augmentation technologies, nanotube electronics, volumetric displays, 4D printing technologies, or others.
The storage systems described above may also, either alone or in combination with other computing environments, be used to support a wide variety of digital platforms. Such digital platforms can include, for example, 5G wireless systems and platforms, digital twin platforms, edge computing platforms, IoT platforms, quantum computing platforms, serverless PaaS, software-defined security, neuromorphic computing platforms, and so on.
The storage systems described above may also be part of a multi-cloud environment in which multiple cloud computing and storage services are deployed in a single heterogeneous architecture. In order to facilitate the operation of such a multi-cloud environment, DevOps tools may be deployed to enable orchestration across clouds. Likewise, continuous development and continuous integration tools may be deployed to standardize processes around continuous integration and delivery, new feature rollout and provisioning cloud workloads. By standardizing these processes, a multi-cloud strategy may be implemented that enables the utilization of the best provider for each workload.
The storage systems described above may be used as a part of a platform to enable the use of crypto-anchors that may be used to authenticate a product's origins and contents to ensure that it matches a blockchain record associated with the product. Similarly, as part of a suite of tools to secure data stored on the storage system, the storage systems described above may implement various encryption technologies and schemes, including lattice cryptography. Lattice cryptography can involve constructions of cryptographic primitives that involve lattices, either in the construction itself or in the security proof. Unlike public-key schemes such as the RSA, Diffie-Hellman or Elliptic-Curve cryptosystems, which are easily attacked by a quantum computer, some lattice-based constructions appear to be resistant to attack by both classical and quantum computers.
A quantum computer is a device that performs quantum computing. Quantum computing is computing using quantum-mechanical phenomena, such as superposition and entanglement. Quantum computers differ from traditional computers that are based on transistors, as such traditional computers require that data be encoded into binary digits (bits), each of which is always in one of two definite states (0 or 1). In contrast to traditional computers, quantum computers use quantum bits, which can be in superpositions of states. A quantum computer maintains a sequence of qubits, where a single qubit can represent a one, a zero, or any quantum superposition of those two qubit states. A pair of qubits can be in any quantum superposition of 4 states, and three qubits in any superposition of 8 states. A quantum computer with n qubits can generally be in an arbitrary superposition of up to 2{circumflex over ( )}n different states simultaneously, whereas a traditional computer can only be in one of these states at any one time. A quantum Turing machine is a theoretical model of such a computer.
The storage systems described above may also be paired with FPGA-accelerated servers as part of a larger AI or ML infrastructure. Such FPGA-accelerated servers may reside near (e.g., in the same data center) the storage systems described above or even incorporated into an appliance that includes one or more storage systems, one or more FPGA-accelerated servers, networking infrastructure that supports communications between the one or more storage systems and the one or more FPGA-accelerated servers, as well as other hardware and software components. Alternatively, FPGA-accelerated servers may reside within a cloud computing environment that may be used to perform compute-related tasks for AI and ML jobs. Any of the embodiments described above may be used to collectively serve as a FPGA-based AI or ML platform. Readers will appreciate that, in some embodiments of the FPGA-based AI or ML platform, the FPGAs that are contained within the FPGA-accelerated servers may be reconfigured for different types of ML models (e.g., LSTMs, CNNs, GRUs). The ability to reconfigure the FPGAs that are contained within the FPGA-accelerated servers may enable the acceleration of a ML or AI application based on the most optimal numerical precision and memory model being used. Readers will appreciate that by treating the collection of FPGA-accelerated servers as a pool of FPGAs, any CPU in the data center may utilize the pool of FPGAs as a shared hardware microservice, rather than limiting a server to dedicated accelerators plugged into it.
The FPGA-accelerated servers and the GPU-accelerated servers described above may implement a model of computing where, rather than keeping a small amount of data in a CPU and running a long stream of instructions over it as occurred in more traditional computing models, the machine learning model and parameters are pinned into the high-bandwidth on-chip memory with lots of data streaming through the high-bandwidth on-chip memory. FPGAs may even be more efficient than GPUs for this computing model, as the FPGAs can be programmed with only the instructions needed to run this kind of computing model.
The storage systems described above may be configured to provide parallel storage, for example, through the use of a parallel file system such as BeeGFS. Such parallel files systems may include a distributed metadata architecture. For example, the parallel file system may include a plurality of metadata servers across which metadata is distributed, as well as components that include services for clients and storage servers.
The systems described above can support the execution of a wide array of software applications. Such software applications can be deployed in a variety of ways, including container-based deployment models. Containerized applications may be managed using a variety of tools. For example, containerized applications may be managed using Docker Swarm, Kubernetes, and others. Containerized applications may be used to facilitate a serverless, cloud native computing deployment and management model for software applications. In support of a serverless, cloud native computing deployment and management model for software applications, containers may be used as part of an event handling mechanisms (e.g., AWS Lambdas) such that various events cause a containerized application to be spun up to operate as an event handler.
The systems described above may be deployed in a variety of ways, including being deployed in ways that support fifth generation (‘5G’) networks. 5G networks may support substantially faster data communications than previous generations of mobile communications networks and, as a consequence may lead to the disaggregation of data and computing resources as modern massive data centers may become less prominent and may be replaced, for example, by more-local, micro data centers that are close to the mobile-network towers. The systems described above may be included in such local, micro data centers and may be part of or paired to multi-access edge computing (‘MEC’) systems. Such MEC systems may enable cloud computing capabilities and an IT service environment at the edge of the cellular network. By running applications and performing related processing tasks closer to the cellular customer, network congestion may be reduced and applications may perform better.
The storage systems described above may also be configured to implement NVMe Zoned Namespaces. Through the use of NVMe Zoned Namespaces, the logical address space of a namespace is divided into zones. Each zone provides a logical block address range that must be written sequentially and explicitly reset before rewriting, thereby enabling the creation of namespaces that expose the natural boundaries of the device and offload management of internal mapping tables to the host. In order to implement NVMe Zoned Name Spaces (‘ZNS’), ZNS SSDs or some other form of zoned block devices may be utilized that expose a namespace logical address space using zones. With the zones aligned to the internal physical properties of the device, several inefficiencies in the placement of data can be eliminated. In such embodiments, each zone may be mapped, for example, to a separate application such that functions like wear levelling and garbage collection could be performed on a per-zone or per-application basis rather than across the entire device. In order to support ZNS, the storage controllers described herein may be configured with to interact with zoned block devices through the usage of, for example, the Linux™ kernel zoned block device interface or other tools.
The storage systems described above may also be configured to implement zoned storage in other ways such as, for example, through the usage of shingled magnetic recording (SMR) storage devices. In examples where zoned storage is used, device-managed embodiments may be deployed where the storage devices hide this complexity by managing it in the firmware, presenting an interface like any other storage device. Alternatively, zoned storage may be implemented via a host-managed embodiment that depends on the operating system to know how to handle the drive, and only write sequentially to certain regions of the drive. Zoned storage may similarly be implemented using a host-aware embodiment in which a combination of a drive managed and host managed implementation is deployed.
For further explanation,
Communication interface 352 may be configured to communicate with one or more computing devices. Examples of communication interface 352 include, without limitation, a wired network interface (such as a network interface card), a wireless network interface (such as a wireless network interface card), a modem, an audio/video connection, and any other suitable interface.
Processor 354 generally represents any type or form of processing unit capable of processing data and/or interpreting, executing, and/or directing execution of one or more of the instructions, processes, and/or operations described herein. Processor 354 may perform operations by executing computer-executable instructions 362 (e.g., an application, software, code, and/or other executable data instance) stored in storage device 356.
Storage device 356 may include one or more data storage media, devices, or configurations and may employ any type, form, and combination of data storage media and/or device. For example, storage device 356 may include, but is not limited to, any combination of the non-volatile media and/or volatile media described herein. Electronic data, including data described herein, may be temporarily and/or permanently stored in storage device 356. For example, data representative of computer-executable instructions 362 configured to direct processor 354 to perform any of the operations described herein may be stored within storage device 356. In some examples, data may be arranged in one or more databases residing within storage device 356.
I/O module 358 may include one or more I/O modules configured to receive user input and provide user output. I/O module 358 may include any hardware, firmware, software, or combination thereof supportive of input and output capabilities. For example, I/O module 358 may include hardware and/or software for capturing user input, including, but not limited to, a keyboard or keypad, a touchscreen component (e.g., touchscreen display), a receiver (e.g., an RF or infrared receiver), motion sensors, and/or one or more input buttons.
I/O module 358 may include one or more devices for presenting output to a user, including, but not limited to, a graphics engine, a display (e.g., a display screen), one or more output drivers (e.g., display drivers), one or more audio speakers, and one or more audio drivers. In certain embodiments, I/O module 358 is configured to provide graphical data to a display for presentation to a user. The graphical data may be representative of one or more graphical user interfaces and/or any other graphical content as may serve a particular implementation. In some examples, any of the systems, computing devices, and/or other components described herein may be implemented by computing device 350.
The storage systems described above may, either alone or in combination, by configured to serve as a continuous data protection store. A continuous data protection store is a feature of a storage system that records updates to a dataset in such a way that consistent images of prior contents of the dataset can be accessed with a low time granularity (often on the order of seconds, or even less), and stretching back for a reasonable period of time (often hours or days). These allow access to very recent consistent points in time for the dataset, and also allow access to points in time for a dataset that might have just preceded some event that, for example, caused parts of the dataset to be corrupted or otherwise lost, while retaining close to the maximum number of updates that preceded that event. Conceptually, they are like a sequence of snapshots of a dataset taken very frequently and kept for a long period of time, though continuous data protection stores are often implemented quite differently from snapshots. A storage system implementing a data continuous data protection store may further provide a means of accessing these points in time, accessing one or more of these points in time as snapshots or as cloned copies, or reverting the dataset back to one of those recorded points in time.
Over time, to reduce overhead, some points in the time held in a continuous data protection store can be merged with other nearby points in time, essentially deleting some of these points in time from the store. This can reduce the capacity needed to store updates. It may also be possible to convert a limited number of these points in time into longer duration snapshots. For example, such a store might keep a low granularity sequence of points in time stretching back a few hours from the present, with some points in time merged or deleted to reduce overhead for up to an additional day. Stretching back in the past further than that, some of these points in time could be converted to snapshots representing consistent point-in-time images from only every few hours.
In order to identify data segments 426 with similarity of content for garbage collection and data compression together, a segment selector 408 cooperates with a hash function 410 and informs data compression module 414 of selected data segments 426. The garbage collection module 402 computes hashes of portions of selected data segments 426, forming sets of hashes 412, e.g., one set of hashes or hash values per data segment (see
Although a storage system could conceivably directly compare data of two segments, it should be appreciated that this would be computationally expensive. Practical embodiments of a storage system, for the sake of computational efficiency, compare a deterministic subset of hash values of portions of data of two or more segments in application of a similarity metric 510, such as the Jaccard distance 512. It should be appreciated that other similarity metrics besides the Jaccard distance could be devised for further embodiments, in keeping with the teachings herein.
Continuing with
Continuing with
In an action 902, the system obtains hash results from data segments. The data segments are stored in storage memory of the storage system. For example, the system could use a sliding window hash function applied at offsets in the data segment as described above, and deterministically select a subset of hash values for each data segment, for comparison with subsets of hash values for other data segments. In an action 904, the system determines similarity of content of data segments, based on the hash results. For example, the system could use a similarity metric and a similarity threshold as described above in comparing the hash results across data segments, and could select data segments for compression together when two or more data segments have a similarity of content that meets the similarity threshold.
In an action 906, the system compresses live data of two or more data segments that have similarity of content meeting the similarity threshold. For example, the system could compress live data using data compression that recognizes identical and/or similar portions of data, recording differences from a reference data for similar data as described above. Such data segments with similarity of content are compressed together, for data compression and data storage efficiency.
In an action 908 of
Although some embodiments are described largely in the context of a storage system, readers of skill in the art will recognize that embodiments of the present disclosure may also take the form of a computer program product disposed upon computer readable storage media for use with any suitable processing system. Such computer readable storage media may be any storage medium for machine-readable information, including magnetic media, optical media, solid-state media, or other suitable media. Examples of such media include magnetic disks in hard drives or diskettes, compact disks for optical drives, magnetic tape, and others as will occur to those of skill in the art. Persons skilled in the art will immediately recognize that any computer system having suitable programming means will be capable of executing the steps described herein as embodied in a computer program product. Persons skilled in the art will recognize also that, although some of the embodiments described in this specification are oriented to software installed and executing on computer hardware, nevertheless, alternative embodiments implemented as firmware or as hardware are well within the scope of the present disclosure.
In some examples, a non-transitory computer-readable medium storing computer-readable instructions may be provided in accordance with the principles described herein. The instructions, when executed by a processor of a computing device, may direct the processor and/or computing device to perform one or more operations, including one or more of the operations described herein. Such instructions may be stored and/or transmitted using any of a variety of known computer-readable media.
A non-transitory computer-readable medium as referred to herein may include any non-transitory storage medium that participates in providing data (e.g., instructions) that may be read and/or executed by a computing device (e.g., by a processor of a computing device). For example, a non-transitory computer-readable medium may include, but is not limited to, any combination of non-volatile storage media and/or volatile storage media. Exemplary non-volatile storage media include, but are not limited to, read-only memory, flash memory, a solid-state drive, a magnetic storage device (e.g. a hard disk, a floppy disk, magnetic tape, etc.), ferroelectric random-access memory (“RAM”), and an optical disc (e.g., a compact disc, a digital video disc, a Blu-ray disc, etc.). Exemplary volatile storage media include, but are not limited to, RAM (e.g., dynamic RAM).
Advantages and features of the present disclosure can be further described by the following statements:
One or more embodiments may be described herein with the aid of method steps illustrating the performance of specified functions and relationships thereof. The boundaries and sequence of these functional building blocks and method steps have been arbitrarily defined herein for convenience of description. Alternate boundaries and sequences can be defined so long as the specified functions and relationships are appropriately performed. Any such alternate boundaries or sequences are thus within the scope and spirit of the claims. Further, the boundaries of these functional building blocks have been arbitrarily defined for convenience of description. Alternate boundaries could be defined as long as the certain significant functions are appropriately performed. Similarly, flow diagram blocks may also have been arbitrarily defined herein to illustrate certain significant functionality.
To the extent used, the flow diagram block boundaries and sequence could have been defined otherwise and still perform the certain significant functionality. Such alternate definitions of both functional building blocks and flow diagram blocks and sequences are thus within the scope and spirit of the claims. One of average skill in the art will also recognize that the functional building blocks, and other illustrative blocks, modules and components herein, can be implemented as illustrated or by discrete components, application specific integrated circuits, processors executing appropriate software and the like or any combination thereof.
While particular combinations of various functions and features of the one or more embodiments are expressly described herein, other combinations of these features and functions are likewise possible. The present disclosure is not limited by the particular examples disclosed herein and expressly incorporates these other combinations.
This application is a continuation application for patent entitled to a filing date and claiming the benefit of earlier-filed U.S. patent application Ser. No. 17/156,979, filed Jan. 25, 2021, which is hereby incorporated herein by reference in its entirety.
| Number | Date | Country | |
|---|---|---|---|
| Parent | 17156979 | Jan 2021 | US |
| Child | 18799925 | US |
