Rate matching technique for balancing segment cleaning and I/O workload

Abstract

A rate matching technique may be configured to adjust a rate of cleaning of one or more selected segments of the storage array to accommodate a variable rate of incoming workload processed by a storage input/output (I/O) stack executing on one or more nodes of a cluster. An extent store layer of the storage I/O stack may clean a segment in accordance with segment cleaning which, illustratively, may be embodied as a segment cleaning process. The rate matching technique may be implemented as a feedback control mechanism configured to adjust the segment cleaning process based on the incoming workload. Components of the feedback control mechanism may include one or more weight schedulers and various accounting data structures, e.g., counters, configured to track the progress of segment cleaning and free space usage. The counters may also be used to balance the rates of segment cleaning and incoming I/O workload, which may change depending upon an incoming I/O rate. When the incoming I/O rate changes, the rate of segment cleaning may be adjusted accordingly to ensure that rates are substantially balanced.

Description

BACKGROUND

Technical Field

The present disclosure relates to storage systems and, more specifically, to efficient segment cleaning in a storage system.

Background Information

A storage system typically includes one or more storage devices, such as solid state drives (SSDs) embodied as flash storage devices of a storage array, into which information may be entered, and from which the information may be obtained, as desired. The storage system may implement a high-level module, such as a file system, to logically organize the information stored on the storage devices of the array as storage containers, such as files or logical units (LUNs). The storage system cluster may be configured to operate according to a client/server model of information delivery to thereby allow one or more clients (hosts) to access the storage containers, e.g., via an I/O request. Each storage container may be implemented as a set of data structures, such as data blocks that store data for the storage containers and metadata blocks that describe the data of the storage containers. For example, the metadata may describe, e.g., identify, storage locations on the devices for the data.

Some types of SSDs, especially those with NAND flash components, may or may not include an internal controller (i.e., inaccessible to a user of the SSD) that moves valid data from old locations to new locations among those components at the granularity of a page (e.g., 8 Kbytes) and then only to previously-erased pages. Thereafter, the old locations where the pages were stored are freed and available to store additional data (e.g., received via the I/O request). Moving of valid data from old to new locations. i.e., garbage collection, contributes to write amplification in the system. Although needed to provide available locations on the SSDs of the storage array (i.e., storage space) to write the additional data, such garbage collection should be performed so as to maintain smooth latency from a host perspective (i.e., maintain a bounded latency for the I/O request). It is thus desirable to continue servicing data (i.e., processing I/O requests) at the storage system within the bounded latency, while ensuring there is sufficient storage space and bandwidth to garbage collect and write the data to the free locations of the storage array.

BRIEF DESCRIPTION OF THE DRAWINGS

The above and further advantages of the embodiments herein may be better understood by referring to the following description in conjunction with the accompanying drawings in which like reference numerals indicate identically or functionally similar elements, of which:

FIG. 1 is a block diagram of a plurality of nodes interconnected as a cluster;

FIG. 2 is a block diagram of a node;

FIG. 3 is a block diagram of a storage input/output (I/O) stack of the node;

FIG. 4 illustrates a write path of the storage I/O stack;

FIG. 5 illustrates a read path of the storage I/O stack;

FIG. 6 illustrates segment cleaning by a layered file system of the storage I/O stack;

FIG. 7 illustrates a RAID stripe formed by the layered file system;

FIGS. 8A and 8B illustrate read and write weight schedules respectively of a rate matching technique; and

FIG. 9 illustrates the rate matching technique.

OVERVIEW

The embodiments described herein are directed to a rate matching technique configured to adjust a rate of cleaning of one or more selected portions or segments of a storage array to accommodate a variable rate of incoming workload processed by a storage input/output (I/O) stack executing on one or more nodes of a cluster. The incoming workload may manifest as I/O operations, e.g., read and write operations, directed to user data, e.g., from a host coupled to the cluster, and associated metadata, e.g., from a volume layer and an extent store layer of the storage I/O stack. The extent store layer provides sequential storage of the user data and metadata on solid state drives (SSDs) of the storage array. The user data (and metadata) may be organized as an arbitrary number of variable-length extents of one or more host-visible logical units (LUNs) served by the nodes. The metadata may include mappings from host-visible logical block address ranges (i.e., offset ranges) of a LUN to extent keys, as well as mappings of the extent keys to locations of the extents stored on the SSDs. The extent store layer may clean a segment in accordance with segment cleaning (i.e., garbage collection) which, illustratively, may be embodied as a segment cleaning process. The segment cleaning process may read all valid extents from one or more segments to be cleaned and write those valid extents (e.g., extents not deleted or overwritten) to one or more other segments available to be written, to thereby free-up (i.e., “clean”) storage space of the segments being cleaned.

In an embodiment, the rate matching technique may be implemented as a feedback control mechanism (e.g., a feedback control loop) configured to adjust the segment cleaning process based on the incoming read and write workload. Components of the feedback control mechanism may include one or more weight schedulers and various accounting data structures, e.g., counters, configured to track (determine) the progress, e.g., rate, of segment cleaning and free space usage. The counters may also be used to balance the rates of segment cleaning and incoming I/O workload, which may change depending upon an incoming I/O rate and a pattern of overwrites in the incoming I/O workload (reducing cleaning). When the incoming I/O rate changes, the rate of segment cleaning may be adjusted accordingly to ensure that rates (i.e., incoming I/O rate and segment cleaning rate) are substantially the same, i.e., balanced. In this manner, segment cleaning is performed only when needed (i.e., freeing space for the incoming I/O) and reduces write amplification.

The I/O operations (e.g., read and write operations) in the storage I/O stack may be categorized as user, metadata, and relocation (segment cleaner) I/O operations. Read operations are serviced per SSD and directly dispatched to a RAID layer of the storage I/O stack from a reader process of the extent store layer. Write operations are accumulated at a writer process of the extent store layer, where associated extents may be packed to form a full stripe before being dispatched to the RAID layer. The weight schedulers may be employed at the extent store layer to regulate the read operations and full stripe write operations at the RAID layer. Illustratively, a write weight scheduler is provided for all SSDs of a segment at the writer process, whereas a read weight scheduler is provided for every SSD of the segment at the reader process in order to control bandwidth distribution between the various categories of I/O operations.

In an embodiment, each weight scheduler may include a plurality of queues, wherein each queue is associated with a category of I/O operations, e.g., user, metadata and relocation. That is, the queues may include (i) a data queue for incoming user data workload, (ii) a metadata queue for incoming volume and extent store layer metadata workload, and (iii) a relocation queue for data and metadata relocation during segment cleaning. According to the rate matching technique, the weight schedulers may be configured to exert fine-grain control (a secondary control loop) over the rate of segment cleaning to match the rate of the incoming I/O workload by assigning relative weights to the queues to thereby ensure that the segment cleaning does not introduce unnecessary write amplification by operating too quickly and, in addition, does not impact performance that may increase a host-perceived latency associated with the I/O workload by operating either too quickly or too slowly.

DESCRIPTION

Storage Cluster

FIG. 1 is a block diagram of a plurality of nodes 200 interconnected as a cluster 100 and configured to provide storage service relating to the organization of information on storage devices. The nodes 200 may be interconnected by a cluster interconnect fabric 110 and include functional components that cooperate to provide a distributed storage architecture of the cluster 100, which may be deployed in a storage area network (SAN). As described herein, the components of each node 200 include hardware and software functionality that enable the node to connect to one or more hosts 120 over a computer network 130, as well as to one or more storage arrays 150 of storage devices over a storage interconnect 140, to thereby render the storage service in accordance with the distributed storage architecture.

Each host 120 may be embodied as a general-purpose computer configured to interact with any node 200 in accordance with a client/server model of information delivery. That is, the client (host) may request the services of the node, and the node may return the results of the services requested by the host, by exchanging packets over the network 130. The host may issue packets including file-based access protocols, such as the Network File System (NFS) protocol over the Transmission Control Protocol/Internet Protocol (TCP/IP), when accessing information on the node in the form of storage containers such as files and directories. However, in an embodiment, the host 120 illustratively issues packets including block-based access protocols, such as the Small Computer Systems Interface (SCSI) protocol encapsulated over TCP (iSCSI) and SCSI encapsulated over FC, when accessing information in the form of storage containers such as logical units (LUNs). Notably, any of the nodes 200 may service a request directed to a storage container stored on the cluster 100.

FIG. 2 is a block diagram of a node 200 that is illustratively embodied as a storage system having one or more central processing units (CPUs) 210 coupled to a memory 220 via a memory bus 215. The CPU 210 is also coupled to a network adapter 230, storage controllers 240, a cluster interconnect interface 250, and a non-volatile random access memory (NVRAM 280) via a system interconnect 270. The network adapter 230 may include one or more ports adapted to couple the node 200 to the host(s) 120 over computer network 130, which may include point-to-point links, wide area networks, virtual private networks implemented over a public network (Internet) or a local area network. The network adapter 230 thus includes the mechanical, electrical and signaling circuitry needed to connect the node to the network 130, which illustratively embodies an Ethernet or Fibre Channel (FC) network.

The memory 220 may include memory locations that are addressable by the CPU 210 for storing software programs and data structures associated with the embodiments described herein. The CPU 210 may, in turn, include processing elements and/or logic circuitry configured to execute the software programs, such as a storage input/output (I/O) stack 300, and manipulate the data structures. Illustratively, the storage I/O stack 300 may be implemented as a set of user mode processes that may be decomposed into a plurality of threads. An operating system kernel 224, portions of which are typically resident in memory 220 (in-core) and executed by the processing elements (i.e., CPU 210), functionally organizes the node by, inter alia, invoking operations in support of the storage service implemented by the node and, in particular, the storage I/O stack 300. A suitable operating system kernel 224 may include a general-purpose operating system, such as the UNIX® series or Microsoft Windows® series of operating systems, or an operating system with configurable functionality such as microkernels and embedded kernels. However, in an embodiment described herein, the operating system kernel is illustratively the Linux® operating system. It will be apparent to those skilled in the art that other processing and memory means, including various computer readable media, may be used to store and execute program instructions pertaining to the embodiments herein.

Each storage controller 240 cooperates with the storage I/O stack 300 executing on the node 200 to access information requested by the host 120. The information is preferably stored on storage devices such as solid state drives (SSDs) 260, illustratively embodied as flash storage devices, of storage array 150. In an embodiment, the flash storage devices may be based on NAND flash components, e.g., single-layer-cell (SLC) flash, multi-layer-cell (MLC) flash or triple-layer-cell (TLC) flash, although it will be understood to those skilled in the art that other non-volatile, solid-state electronic devices (e.g., drives based on storage class memory components) may be advantageously used with the embodiments described herein. Accordingly, the storage devices may or may not be block-oriented (i.e., accessed as blocks). The storage controller 240 includes one or more ports having I/O interface circuitry that couples to the SSDs 260 over the storage interconnect 140, illustratively embodied as a serial attached SCSI (SAS) topology. Alternatively, other point-to-point I/O interconnect arrangements may be used, such as a conventional serial ATA (SATA) topology or a PCI topology. The system interconnect 270 may also couple the node 200 to a local service storage device 248, such as an SSD, configured to locally store cluster-related configuration information, e.g., as cluster database (DB) 244, which may be replicated to the other nodes 200 in the cluster 100.

The cluster interconnect interface 250 may include one or more ports adapted to couple the node 200 to the other node(s) of the cluster 100. In an embodiment, Ethernet may be used as the clustering protocol and interconnect fabric media, although it will be apparent to those skilled in the art that other types of protocols and interconnects, such as Infiniband, may be utilized within the embodiments described herein. The NVRAM 280 may include a back-up battery or other built-in last-state retention capability (e.g., non-volatile semiconductor memory such as storage class memory) that is capable of maintaining data in light of a failure to the node and cluster environment. Illustratively, a portion of the NVRAM 280 may be configured as one or more non-volatile logs (NV Logs 285) configured to temporarily record (“log”) I/O requests, such as write requests, received from the host 120.

Storage I/O Stack

FIG. 3 is a block diagram of the storage I/O stack 300 that may be advantageously used with one or more embodiments described herein. The storage I/O stack 300 includes a plurality of software modules or layers that cooperate with other functional components of the nodes 200 to provide the distributed storage architecture of the cluster 100. In an embodiment, the distributed storage architecture presents an abstraction of a single storage container, i.e., all of the storage arrays 150 of the nodes 200 for the entire cluster 100 organized as one large pool of storage. In other words, the architecture consolidates storage, i.e., the SSDs 260 of the arrays 150, throughout the cluster (retrievable via cluster-wide keys) to enable storage of the LUNs. Both storage capacity and performance may then be subsequently scaled by adding nodes 200 to the cluster 100.

Illustratively, the storage I/O stack 300 includes an administration layer 310, a protocol layer 320, a persistence layer 330, a volume layer 340, an extent store layer 350, a Redundant Array of Independent Disks (RAID) layer 360, a storage layer 365 and a NVRAM (storing NV Logs) “layer” interconnected with a messaging kernel 370. The messaging kernel 370 may provide a message-based (or event-based) scheduling model (e.g., asynchronous scheduling) that employs messages as fundamental units of work exchanged (i.e., passed) among the layers. Suitable message-passing mechanisms provided by the messaging kernel to transfer information between the layers of the storage I/O stack 300 may include, e.g., for intra-node communication: i) messages that execute on a pool of threads, ii) messages that execute on a single thread progressing as an operation through the storage I/O stack, iii) messages using an Inter Process Communication (IPC) mechanism, and, e.g., for inter-node communication: messages using a Remote Procedure Call (RPC) mechanism in accordance with a function shipping implementation. Alternatively, the I/O stack may be implemented using a thread-based or stack-based execution model. In one or more embodiments, the messaging kernel 370 allocates processing resources from the operating system kernel 224 to execute the messages. Each storage I/O stack layer may be implemented as one or more instances (i.e., processes) executing one or more threads (e.g., in kernel or user space) that process the messages passed between the layers such that the messages provide synchronization for blocking and non-blocking operation of the layers.

In an embodiment, the protocol layer 320 may communicate with the host 120 over the network 130 by exchanging discrete frames or packets configured as I/O requests according to pre-defined protocols, such as iSCSI and FCP. An I/O request, e.g., a read or write request, may be directed to a LUN and may include I/O parameters such as, inter alia, a LUN identifier (ID), a logical block address (LBA) of the LUN, a length (i.e., amount of data) and, in the case of a write request, write data. The protocol layer 320 receives the I/O request and forwards it to the persistence layer 330, which records the request into a persistent write-back cache 380 illustratively embodied as a log whose contents can be replaced randomly, e.g., under some random access replacement policy rather than only in serial fashion, and returns an acknowledgement to the host 120 via the protocol layer 320. In an embodiment only I/O requests that modify the LUN, e.g., write requests, are logged. Notably, the I/O request may be logged at the node receiving the I/O request, or in an alternative embodiment in accordance with the function shipping implementation, the I/O request may be logged at another node.

Illustratively, dedicated logs may be maintained by the various layers of the storage I/O stack 300. For example, a dedicated log 335 may be maintained by the persistence layer 330 to record the I/O parameters of an I/O request as equivalent internal, i.e., storage I/O stack, parameters, e.g., volume ID, offset, and length. In the case of a write request, the persistence layer 330 may also cooperate with the NVRAM 280 to implement the write-back cache 380 configured to store the write data associated with the write request. In an embodiment, the write-back cache may be structured as a log. Notably, the write data for the write request may be physically stored in the cache 380 such that the log 335 contains the reference to the associated write data. It will be understood to persons skilled in the art that other variations of data structures may be used to store or maintain the write data in NVRAM including data structures with no logs. In an embodiment, a copy of the write-back cache may be also maintained in the memory 220 to facilitate direct memory access to the storage controllers. In other embodiments, caching may be performed at the host 120 or at a receiving node in accordance with a protocol that maintains coherency between the data stored at the cache and the cluster.

In an embodiment, the administration layer 310 may apportion the LUN into multiple volumes, each of which may be partitioned into multiple regions (e.g., allotted as disjoint block address ranges), with each region having one or more segments stored as multiple stripes on the array 150. A plurality of volumes distributed among the nodes 200 may thus service a single LUN, i.e., each volume within the LUN services a different LBA range (i.e., offset range and length, hereinafter offset range) or set of ranges within the LUN. Accordingly, the protocol layer 320 may implement a volume mapping technique to identify a volume to which the I/O request is directed (i.e., the volume servicing the offset range indicated by the parameters of the I/O request). Illustratively, the cluster database 244 may be configured to maintain one or more associations (e.g., key-value pairs) for each of the multiple volumes, e.g., an association between the LUN ID and a volume, as well as an association between the volume and a node ID for a node managing the volume. The administration layer 310 may also cooperate with the database 244 to create (or delete) one or more volumes associated with the LUN (e.g., creating a volume ID/LUN key-value pair in the database 244). Using the LUN ID and LBA (or LBA range), the volume mapping technique may provide a volume ID (e.g., using appropriate associations in the cluster database 244) that identifies the volume and node servicing the volume destined for the request as well as translate the LBA (or LBA range) into an offset and length within the volume. Specifically, the volume ID is used to determine a volume layer instance that manages volume metadata associated with the LBA or LBA range. As noted, the protocol layer 320 may pass the I/O request (i.e., volume ID, offset and length) to the persistence layer 330, which may use the function shipping (e.g., inter-node) implementation to forward the I/O request to the appropriate volume layer instance executing on a node in the cluster based on the volume ID.

In an embodiment, the volume layer 340 may manage the volume metadata by, e.g., maintaining states of host-visible containers, such as ranges of LUNs, and performing data management functions, such as creation of snapshots and clones, for the LUNs in cooperation with the administration layer 310. The volume metadata is illustratively embodied as in-core mappings from LUN addresses (i.e., offsets) to durable extent keys, which are unique cluster-wide IDs associated with SSD storage locations for extents within an extent key space of the cluster-wide storage container. That is, an extent key may be used to retrieve the data of the extent at an SSD storage location associated with the extent key. Alternatively, there may be multiple storage containers in the cluster wherein each container has its own extent key space, e.g., where the administration layer 310 provides distribution of extents among the storage containers. An extent is a variable length block of data that provides a unit of storage on the SSDs and that need not be aligned on any specific boundary, i.e., it may be byte aligned. Accordingly, an extent may be an aggregation of write data from a plurality of write requests to maintain such alignment. Illustratively, the volume layer 340 may record the forwarded request (e.g., information or parameters characterizing the request), as well as changes to the volume metadata, in dedicated log 345 maintained by the volume layer 340. Subsequently, the contents of the volume layer log 345 may be written to the storage array 150 in accordance with a checkpoint (e.g., synchronization) operation that stores in-core metadata on the array 150. That is, the checkpoint operation (checkpoint) ensures that a consistent state of metadata, as processed in-core, is committed to (i.e., stored on) the storage array 150; whereas the retirement of log entries ensures that the entries accumulated in the volume layer log 345 synchronize with the metadata checkpoints committed to the storage array 150 by, e.g., retiring those accumulated log entries prior to the checkpoint. In one or more embodiments, the checkpoint and retirement of log entries may be data driven, periodic or both.

In an embodiment, the extent store layer 350 is responsible for storing extents on the SSDs 260 (i.e., on the storage array 150) and for providing the extent keys to the volume layer 340 (e.g., in response to a forwarded write request). The extent store layer 350 is also responsible for retrieving data (e.g., an existing extent) using an extent key (e.g., in response to a forwarded read request). The extent store layer 350 may be responsible for performing de-duplication and compression on the extents prior to storage. The extent store layer 350 may maintain in-core mappings (e.g., embodied as hash tables) of extent keys to SSD storage locations (e.g., offset on an SSD 260 of array 150). The extent store layer 350 may also maintain a dedicated log 355 of entries that accumulate requested “put” and “delete” operations (i.e., write requests and delete requests for extents issued from other layers to the extent store layer 350), where these operations change the in-core mappings (i.e., hash table entries). Subsequently, the in-core mappings and contents of the extent store layer log 355 may be written to the storage array 150 in accordance with a “fuzzy” checkpoint 390 (i.e., checkpoints with incremental changes recorded in one or more log files) in which selected in-core mappings (less than the total), are committed to the array 150 at various intervals (e.g., driven by an amount of change to the in-core mappings, size thresholds of log 355, or periodically). Notably, the accumulated entries in log 355 may be retired once all in-core mappings have been committed to include the changes recorded in those entries.

In an embodiment, the RAID layer 360 may organize the SSDs 260 within the storage array 150 as one or more RAID groups (e.g., sets of SSDs) that enhance the reliability and integrity of extent storage on the array by writing data “stripes” having redundant information, i.e., appropriate parity information with respect to the striped data, across a given number of SSDs 260 of each RAID group. The RAID layer 360 may also store a number of stripes (e.g., stripes of sufficient depth), e.g., in accordance with a plurality of contiguous range write operations, so as to reduce data relocation (i.e., internal flash block management) that may occur within the SSDs as a result of the operations. In an embodiment, the storage layer 365 implements storage I/O drivers that may communicate directly with hardware (e.g., the storage controllers and cluster interface) cooperating with the operating system kernel 224, such as a Linux virtual function I/O (VFIO) driver.

Write Path

FIG. 4 illustrates an I/O (e.g., write) path 400 of the storage I/O stack 300 for processing an I/O request, e.g., a SCSI write request 410. The write request 410 may be issued by host 120 and directed to a LUN stored on the storage arrays 150 of the cluster 100. Illustratively, the protocol layer 320 receives and processes the write request by decoding 420 (e.g., parsing and extracting) fields of the request, e.g., LUN ID, LBA and length (shown at 413), as well as write data 414. The protocol layer 320 may use the results 422 from decoding 420 for a volume mapping technique 430 (described above) that translates the LUN ID and LBA range (i.e., equivalent offset and length) of the write request to an appropriate volume layer instance, i.e., volume ID (volume 445), in the cluster 100 that is responsible for managing volume metadata for the LBA range. In an alternative embodiment, the persistence layer 330 may implement the above described volume mapping technique 430. The protocol layer then passes the results 432, e.g., volume ID, offset, length (as well as write data), to the persistence layer 330, which records the request in the persistence layer log 335 and returns an acknowledgement to the host 120 via the protocol layer 320. The persistence layer 330 may aggregate and organize write data 414 from one or more write requests into a new extent 610 and perform a hash computation, i.e., a hash function, on the new extent to generate a hash value 472 in accordance with an extent hashing technique 450.

The persistence layer 330 may then pass the write request with aggregated write data including, e.g., the volume ID, offset and length, as parameters 434 to the appropriate volume layer instance. In an embodiment, message passing of the parameters 434 (received by the persistence layer) may be redirected to another node via the function shipping mechanism, e.g., RPC, for inter-node communication. Alternatively, message passing of the parameters 434 may be via the IPC mechanism, e.g., message threads, for intra-node communication.

In one or more embodiments, a bucket mapping technique 476 is provided that translates the hash value 472 to an instance of an appropriate extent store layer (i.e., extent store instance 470) that is responsible for storing the new extent 610. Note, the bucket mapping technique may be implemented in any layer of the storage I/O stack above the extent store layer. In an embodiment, for example, the bucket mapping technique may be implemented in the persistence layer 330, the volume layer 340, or a layer that manages cluster-wide information, such as a cluster layer (not shown). Accordingly, the persistence layer 330, the volume layer 340, or the cluster layer may contain computer executable instructions executed by the CPU 210 to perform operations that implement the bucket mapping technique 476 described herein. The persistence layer 330 may then pass the hash value 472 and the new extent 610 to the appropriate volume layer instance and onto the appropriate extent store instance via an extent store put operation. The extent hashing technique 450 may embody an approximately uniform hash function to ensure that any random extent to be written may have an approximately equal chance of falling into any extent store instance 470, i.e., hash buckets are distributed across extent store instances of the cluster 100 based on available resources. As a result, the bucket mapping technique 476 provides load-balancing of write operations (and, by symmetry, read operations) across nodes 200 of the cluster, while also leveling flash wear in the SSDs 260 of the cluster.

In response to the put operation, the extent store instance may process the hash value 472 to perform an extent metadata selection technique 460 that (i) selects an appropriate hash table 480 (e.g., hash table 480a) from a set of hash tables (illustratively in-core) within the extent store instance 470, and (ii) extracts a hash table index 462 from the hash value 472 to index into the selected hash table and lookup a table entry having an extent key 475 identifying a storage location 490 on SSD 260 for the extent. Accordingly, the extent store layer 350 contains computer executable instructions executed by the CPU 210 to perform operations that implement the extent metadata selection technique 460 described herein. If a table entry with a matching extent key is found, then the SSD location 490 mapped from the extent key 475 is used to retrieve an existing extent (not shown) from SSD. The existing extent is then compared with the new extent 610 to determine whether their data is identical. If the data is identical, the new extent 610 is already stored on SSD 260 and a de-duplication opportunity (denoted de-duplication 452) exists such that there is no need to write another copy of the data. Accordingly, a reference count in the table entry for the existing extent is incremented and the extent key 475 of the existing extent is passed to the appropriate volume layer instance for storage within an entry (denoted as volume metadata entry 446) of a dense tree metadata structure 444 (e.g., dense tree 444a), such that the extent key 475 is associated an offset range 440 (e.g., offset range 440a) of the volume 445.

However, if the data of the existing extent is not identical to the data of the new extent 610, a collision occurs and a deterministic algorithm is invoked to sequentially generate as many new candidate extent keys (not shown) mapping to the same bucket as needed to either provide de-duplication 452 or to produce an extent key that is not already stored within the extent store instance. Notably, another hash table (e.g. hash table 480n) may be selected by a new candidate extent key in accordance with the extent metadata selection technique 460. In the event that no de-duplication opportunity exists (i.e., the extent is not already stored) the new extent 610 is compressed in accordance with compression technique 454 and passed to the RAID layer 360, which processes the new extent 610 for storage on SSD 260 within one or more stripes 710 of RAID group 466. The extent store instance may cooperate with the RAID layer 360 to identify a storage segment 650 (i.e., a portion of the storage array 150) and a location on SSD 260 within the segment 650 in which to store the new extent 610. Illustratively, the identified storage segment is a segment with a large contiguous free space having, e.g., location 490 on SSD 260b for storing the extent 610.

In an embodiment, the RAID layer 360 then writes the stripes 710 across the RAID group 466, illustratively as one or more full stripe writes 458. The RAID layer 360 may write a series of stripes 710 of sufficient depth to reduce data relocation that may occur within the flash-based SSDs 260 (i.e., flash block management). The extent store instance then (i) loads the SSD location 490 of the new extent 610 into the selected hash table 480n (i.e., as selected by the new candidate extent key), (ii) passes a new extent key (denoted as extent key 475) to the appropriate volume layer instance for storage within an entry (also denoted as volume metadata entry 446) of a dense tree 444 managed by that volume layer instance, and (iii) records a change to extent metadata of the selected hash table in the extent store layer log 355. Illustratively, the volume layer instance selects dense tree 444a spanning an offset range 440a of the volume 445 that encompasses the offset range of the write request. As noted, the volume 445 (e.g., an offset space of the volume) is partitioned into multiple regions (e.g., allotted as disjoint offset ranges); in an embodiment, each region is represented by a dense tree 444. The volume layer instance then inserts the volume metadata entry 446 into the dense tree 444a and records a change corresponding to the volume metadata entry in the volume layer log 345. Accordingly, the I/O (write) request is sufficiently stored on SSD 260 of the cluster.

Read Path

FIG. 5 illustrates an I/O (e.g., read) path 500 of the storage I/O stack 300 for processing an I/O request, e.g., a SCSI read request 510. The read request 510 may be issued by host 120 and received at the protocol layer 320 of a node 200 in the cluster 100. Illustratively, the protocol layer 320 processes the read request by decoding 420 (e.g., parsing and extracting) fields of the request, e.g., LUN ID, LBA, and length (shown at 513), and uses the decoded results 522, e.g., LUN ID, offset, and length, for the volume mapping technique 430. That is, the protocol layer 320 may implement the volume mapping technique 430 (described above) to translate the LUN ID and LBA range (i.e., equivalent offset and length) of the read request to an appropriate volume layer instance, i.e., volume ID (volume 445), in the cluster 100 that is responsible for managing volume metadata for the LBA (i.e., offset) range. The protocol layer then passes the results 532 to the persistence layer 330, which may search the write cache 380 to determine whether some or all of the read request can be serviced from its cached data. If the entire request cannot be serviced from the cached data, the persistence layer 330 may then pass the remaining portion of the request including, e.g., the volume ID, offset and length, as parameters 534 to the appropriate volume layer instance in accordance with the function shipping mechanism, (e.g., RPC, for inter-node communication) or the IPC mechanism, (e.g., message threads, for intra-node communication).

The volume layer instance may process the read request to access a dense tree metadata structure 444 (e.g., dense tree 444a) associated with a region (e.g., offset range 440a) of a volume 445 that encompasses the requested offset range (specified by parameters 534). The volume layer instance may further process the read request to search for (lookup) one or more volume metadata entries 446 of the dense tree 444a to obtain one or more extent keys 475 associated with one or more extents 610 (or portions of extents) within the requested offset range. In an embodiment, each dense tree 444 may be embodied as multiple levels of a search structure with possibly overlapping offset range entries at each level. The various levels of the dense tree may have volume metadata entries 446 for the same offset, in which case, the higher level has the newer entry and is used to service the read request. A top level of the dense tree 444 is illustratively resident in-core and a page cache 448 may be used to access lower levels of the tree. If the requested range or portion thereof is not present in the top level, a metadata page associated with an index entry at the next lower tree level (not shown) is accessed. The metadata page (i.e., in the page cache 448) at the next level is then searched to find any overlapping entries. This process is then iterated until one or more volume metadata entries 446 of a level are found to ensure that the extent key(s) 475 for the entire requested read range are found. If no metadata entries exist for the entire or portions of the requested range, then the missing portion(s) are zero filled.

Once found, each extent key 475 is processed by the volume layer 340 to, e.g., implement the bucket mapping technique 476 that translates the extent key to an appropriate extent store instance 470 responsible for storing the requested extent 610. Note that, in an embodiment, each extent key 475 may be substantially identical to the hash value 472 associated with the extent 610, i.e., the hash value as calculated during the write request for the extent, such that the bucket mapping 476 and extent metadata selection 460 techniques may be used for both write and read path operations. Note also that the extent key 475 may be derived from the hash value 472. The volume layer 340 may then pass the extent key 475 (i.e., the hash value from a previous write request for the extent) to the appropriate extent store instance 470 (via an extent store get operation), which performs an extent key-to-SSD mapping to determine the location on SSD 260 for the extent.

In response to the get operation, the extent store instance may process the extent key 475 (i.e., hash value 472) to perform the extent metadata selection technique 460 that (i) selects an appropriate hash table 480 (e.g., hash table 480a) from a set of hash tables within the extent store instance 470, and (ii) extracts a hash table index 462 from the extent key 475 (i.e., hash value 472) to index into the selected hash table and lookup a table entry having a matching extent key 475 that identifies a storage location 490 on SSD 260 for the extent 610. That is, the SSD location 490 mapped to the extent key 475 may be used to retrieve the existing extent (denoted as extent 610) from SSD 260 (e.g., SSD 260b). The extent store instance then cooperates with the RAID layer 360 to access the extent on SSD 260b and retrieve the data contents in accordance with the read request. Illustratively, the RAID layer 360 may read the extent in accordance with an extent read operation 468 and pass the extent 610 to the extent store instance. The extent store instance may then decompress the extent 610 in accordance with a decompression technique 456, although it will be understood to those skilled in the art that decompression can be performed at any layer of the storage I/O stack 300. The extent 610 may be stored in a buffer (not shown) in memory 220 and a reference to that buffer may be passed back through the layers of the storage I/O stack. The persistence layer may then load the extent into a read cache 580 (or other staging mechanism) and may extract appropriate read data 512 from the read cache 580 for the LBA range of the read request 510. Thereafter, the protocol layer 320 may create a SCSI read response 514, including the read data 512, and return the read response to the host 120.

Layered File System

The embodiments described herein illustratively employ a layered file system of the storage I/O stack. The layered file system includes a flash-optimized, log-structured layer (i.e., extent store layer) of the file system configured to provide sequential storage of data and metadata (i.e., log-structured layout) on the SSDs 260 of the cluster. The data may be organized as an arbitrary number of variable-length extents of one or more host-visible LUNs served by the nodes. The metadata may include mappings from host-visible logical block address ranges (i.e., offset ranges) of a LUN to extent keys, as well as mappings of the extent keys to SSD storage locations of the extents. Illustratively, the volume layer of the layered file system cooperates with the extent store layer to provide a level of indirection that facilitates efficient log-structured layout of extents on the SSDs by the extent store layer.

In an embodiment, functions of the log-structured layer of the file system, such as write allocation and flash device (i.e., SSD) management, are performed and maintained by the extent store layer 350. Write allocation may include gathering of the variable-length extents to form full stripes that may be written to free segments across SSDs of one or more RAID groups. That is, the log-structured layer of the file system writes extents to initially free (i.e., clean) segments as full stripes rather than partial stripes. Flash device management may include segment cleaning to create such free segments that indirectly map to the SSDs via the RAID groups. Accordingly, partial RAID stripe writes are avoided, which results in reduced RAID-related write amplification.

Instead of relying on garbage collection in the SSDs, the storage I/O stack may implement segment cleaning (i.e., garbage collection) in the extent store layer to bypass performance impacts of flash translation layer (FTL) functionality (including garbage collection) in the SSD. In other words, the storage I/O stack allows the log-structured layer of the file system to operate as a data layout engine using segment cleaning to effectively replace a substantial portion of the FTL functionality of the SSD. The extent store layer may thus process random write requests in accordance with segment cleaning (i.e., garbage collection) to predict flash behavior within its FTL functionality. As a result, a log-structured equivalent source of write amplification for the storage I/O stack may be consolidated and managed at the extent store layer. In addition, the log-structured layer of the file system may be employed, in part, to improve write performance from the flash devices of the storage array.

As noted, the log-structured layout of SSDs is realized by sequentially writing extents to clean segments. Thus, the log-structured layout (i.e., sequential storage) employed by the extent store layer inherently supports variable length extents, thereby allowing unrestricted compression of extents prior to storage on SSD and without specific block level (i.e., in SSD blocks) metadata support from the SSDs, such as 520 byte sectors supporting 512 bytes of data and 8 bytes of metadata (e.g., a pointer to another block containing a tail-end of compressed data.) Typically, consumer grade SSDs support sectors as powers of 2 (e.g., 512 bytes), whereas more expensive enterprise grade SSDs may support enhanced sized sectors (e.g., 520 bytes). Accordingly, the extent store layer may operate with lower cost consumer grade SSDs while supporting variable length extents with their concomitant unfettered compression.

Segment Cleaning

FIG. 6 illustrates segment cleaning by the layered file system. In an embodiment, the extent store layer 350 of the layered file system may write extents to an empty or free region or “segment.” Before rewriting that segment again, the extent store layer 350 may clean the segment in accordance with segment cleaning which, illustratively, may be embodied as a segment cleaning process. The segment cleaning process may read all valid extents 610 from an old segment 650a and write those valid extents (i.e., extents not deleted or overwritten 612) to one or more new segments 650b-c, to thereby free-up (i.e., “clean”) the old segment 650a. New extents may then be written sequentially to the old (now clean) segment. The layered file system may maintain a certain amount of reserve space (i.e., free segments) to enable efficient performance of segment cleaning. For example, the layered file system may illustratively maintain a reserve space of free segments equivalent to approximately 7% of storage capacity. The sequential writing of new extents may manifest as full stripe writes 458, such that a single write operation to storage spans all SSDs in a RAID group 466. Write data may be accumulated until a stripe write operation of a minimum depth can be made.

Illustratively, segment cleaning may be performed to free one or more selected segments that indirectly map to SSDs. As used herein, a SSD may be composed of a plurality of segment chunks 620, wherein each chunk is illustratively approximately 2 GB in size. A segment may include a segment chunk 620a-c from each of a plurality of SSDs in a RAID group 466. Thus, for a RAID group having 24 SSDs, wherein the equivalent storage space of 22 SSDs store data (data SSDs) and the equivalent storage space of 2 SSDs store parity (parity SSDs), each segment may include 44 GB of data and 4 GB of parity. The RAID layer may further configure the RAID groups according to one or more RAID implementations, e.g., RAID 1, 4, 5 and/or 6, to thereby provide protection over the SSDs in the event of, e.g., failure to one or more SSDs. Notably, each segment may be associated with a different RAID group and, thus, may have a different RAID configuration, i.e., each RAID group may be configured according to a different RAID implementation. To free-up or clean selected segments, extents of the segments that contain valid data are moved to different clean segments and the selected segments (now clean) are freed for subsequent reuse. Segment cleaning consolidates fragmented free space to improve write efficiency, e.g., to stripes by reducing RAID-related amplification and to underlying flash blocks by reducing performance impacts of the FTL. Once a segment is cleaned and designated freed, data may be written sequentially to that segment. Accounting structures, e.g., free segment maps indicating an amount of segment free space, maintained by the extent store layer for write allocation, may be employed by the segment cleaning process. Notably, selection of a clean segment to receive data (i.e., writes) from a segment being cleaned may be based upon the amount of free space remaining in the clean segment and/or the last time the clean segment was used. Note further that different portions of data from the segment being cleaned may be moved to different “target” segments. That is, a plurality of relatively clean segments 650b,c may receive differing portions of data from the segment 650a being cleaned.

Illustratively, segment cleaning may cause some write amplification in the storage array (SSDs). However, the file system may reduce such write amplification by writing extents to the SSDs sequentially as a log device. For example, given SSDs with an erase block size of approximately 2 MBs, by writing at least 2 MB of data (extents) sequentially to a free segment, an entire erase block may be overwritten and fragmentation at the SSD level may be eliminated (i.e., reducing garbage collection in the SSD). Yet, the SSDs typically stripe data across multiple flash components and across multiple channels (i.e., storage controllers 240) in order to realize performance. Thus, a relatively large (e.g., 2 GB) write granularity to a free (i.e., clean) segment may be necessary to avoid write amplification at the SSD level (i.e., to override internal SSD striping).

Specifically because the erase block boundaries in the SSD may be unknown, the write granularity should be large enough so that a sequence of writes for extents over a large contiguous range may overwrite previously written extents on the SSD and effectively override garbage collection in the SSDs. In other words, such garbage collection may be preempted because the new data is written over the same range as previous data such that the new data completely overwrites the previously written data. This approach also avoids consuming the reserve space capacity with the new write data. Accordingly, an advantage of the log-structured feature of the storage I/O stack (i.e., log-structured layer of the file system) is the ability to reduce write amplification of the SSDs with only a minimum amount of reserve space in the SSDs. This log-structured feature effectively “moves” flash device management of reserve space from the SSD to the extent store layer, which uses that reserve space to manage the write amplification. Thus, instead of having two sources of write amplification (i.e., the extent store layer and the SSD FTL, which multiply) there is only one source of write amplification (i.e., the extent store layer).

Write Allocation

In an embodiment, there may be multiple RAID stripes per segment. Each time a segment is allocated, i.e., after cleaning the segment, the chunks of various SSDs within the segment may include a series of RAID stripes. The chunks may be at the same or different offsets within the SSDs. The extent store layer may read the chunks sequentially for cleaning purposes and relocate all the valid data to another segment. Thereafter, the chunks 620 of the cleaned segment may be freed and a decision may be rendered as to how to constitute the next segment that uses the chunks. For example, if a SSD is removed from a RAID group, a portion (i.e., a set of chunks 620) of capacity may be omitted from the next segment (i.e., change in RAID stripe configuration) so as to constitute the RAID group from a plurality of chunks 620 that is one chunk narrower, i.e., making the RAID width one less. Thus, by using segment cleaning, a RAID group of the chunks 620 constituting the segments may be effectively created each time a new segment is allocated, i.e., a RAID group is created dynamically from available SSDs when a new segment is allocated. There is generally no requirement to include all of the SSDs 260 in the storage array 150 in the new segment. Alternatively, a chunk 620 from a newly introduced SSD can be added into a RAID group created when a new segment 650 is allocated.

FIG. 7 illustrates a RAID stripe formed by the layered file system. As noted, write allocation may include gathering of the variable-length extents to form one or more stripes across SSDs of one or more RAID groups. In an embodiment, the RAID layer 360 may manage parity computations and topology information used for placement of the extents 610 on the SSDs 260a-n of the RAID group 466. To that end, the RAID layer may cooperate with the extent store layer to organize the extents as stripes 710 within the RAID group. Illustratively, the extent store layer may gather the extents 610 to form one or more full stripes 710 that may be written to a free segment 650a such that a single stripe write operation 458 may span all SSDs in that RAID group. The extent store layer may also cooperate with the RAID layer to pack each stripe 710 as a full stripe of variable-length extents 610. Once the stripe is complete, the RAID layer may pass the full stripe 710 of extents as a set of chunks 620d-f to the storage layer 365 of the storage I/O stack for storage on the SSDs 260. By writing a full stripe (i.e., data and parity) to the free segment, the layered file system avoids the cost of parity updates and spreads any required read operation load across the SSDs. Notably, the extents 610 pending a write operation on an SSD 260 may be accumulated into a chunk 620d,e, which is written as one or more temporally proximate write operations to the SSD (e.g., as 2 Gbytes), thereby reducing the performance impact of the FTL in the SSD.

In an embodiment, an extent store may be viewed as a global pool of extents stored on the storage arrays 150 of the cluster, where each extent may be maintained within a RAID group 466 of an extent store instance. Assume one or more variable-length (i.e., small and/or large) extents are written to a segment. The extent store layer may gather the variable-length extents to form one or more stripes across the SSDs of the RAID group. Although each stripe may include multiple extents 610 and an extent 610c could span more than one stripe 710a,b, each extent is entirely stored on one SSD. In an embodiment, a stripe may have a depth of 16 KB and an extent may have a size of 4 KB, but the extent may thereafter be compressed down to 1 or 2 KB or smaller permitting a larger extent to be packed which may exceed the stripe depth (i.e., the chunk 620g depth). Thus, a stripe may constitute only part of the extent, so the depth of the stripe 710 (i.e., the set of chunks 620d-f constituting the stripe) may be independent of the extent(s) written to any one SSD. Since the extent store layer may write the extents as full stripes across one or more free segments of the SSDs, write amplification associated with processing information of the stripes may be reduced.

Rate Matching Technique

The embodiments described herein are directed to a rate matching technique configured to adjust a rate of cleaning of one or more selected segments of the storage array to accommodate a variable rate of incoming workload (as well as valid data being relocated by cleaning on each selected segment) processed by the storage I/O stack. The incoming workload may manifest as I/O operations, e.g., read and write operations, directed to user data, e.g., from a host coupled to the cluster, and associated metadata, e.g., from the volume layer and extent store layer of the storage I/O stack. As noted, the extent store layer provides sequential storage of the user data and metadata, embodied as extents, on SSDs of the storage array. The extent store layer may clean a segment in accordance with segment cleaning (i.e., garbage collection) which, illustratively, may be embodied as the segment cleaning process. The segment cleaning process may read all valid extents from one or more segments to be cleaned and write those valid extents (e.g., extents not deleted) to one or more other segments available to be written, to thereby free-up (i.e., “clean”) storage space of the segments being cleaned.

A challenge of segment cleaning involves balancing the resources (e.g., I/O bandwidth and CPU) consumed by copying the valid data to one or more other segments with the resources needed to service a variable incoming I/O workload, i.e., manifested as I/O operations. The extent store layer may defer segment cleaning until a number of free segments in the storage array falls below a certain threshold to thereby minimize data and metadata (i.e., extent) relocation and write amplification. That is, segments laden with data need not be cleaned until the number of free segments available for use falls below the threshold. Accordingly, it is desirable to clean a segment fast enough to accommodate the variable incoming write workload, but not too fast so as to clean the segment sooner than is needed and produce needless write amplification, i.e., copying (cleaning) data that may be soon overwritten by the incoming workload. It is possible that extents may be overwritten or deleted prior to cleaning of the segment, in which case there would be no need to relocate the extents to another segment. For example, if a valid extent of a block was relocated to another segment because the segment cleaning process cleaned the segment too fast and then the extent was overwritten (i.e., prior to when the cleaning of the segment was needed), unnecessary write amplification would result. In addition, it is desirable to segment clean as smooth as possible, i.e., maintaining a bounded latency in response to the incoming I/O workload, to avoid an uneven (unbalanced) response time to the host. To that end, the rate matching technique described herein is directed to adjusting segment cleaning to accommodate the variable workload and smooth host-perceived latency.

In an embodiment, the rate matching technique may be implemented as a feedback control mechanism (e.g., a feedback control loop) configured to adjust the segment cleaning process based on the incoming read and write workload and the rate of segment cleaning. Components of the feedback control mechanism may include one or more weight schedulers and various accounting data structures, e.g., counters, configured to track (determine) the progress, e.g., rate, of segment cleaning and free space usage within a segment. The counters may also be used to balance the rates of segment cleaning and incoming I/O workload, which may change depending upon an incoming I/O rate and a pattern of overwrites in the incoming I/O workload (reducing cleaning). When the incoming I/O rate changes, the rate of segment cleaning may be adjusted accordingly to ensure that rates (i.e., incoming I/O rate and segment cleaning rate) are substantially the same, i.e., balanced. In this manner, segment cleaning is performed only when needed (i.e., freeing space for the incoming I/O) and reduces write amplification. Similarly, the incoming I/O rate may be inhibited (i.e., restrained) to allow segment cleaning in retard to catch-up.

FIGS. 8A and 8B illustrate read and write weight schedulers respectively of the rate matching technique. The I/O operations (e.g., read and write operations) in the storage I/O stack may be categorized as user, metadata, and relocation (segment cleaner) I/O operations. Read operations are serviced per SSD 260a-n and directly dispatched to the RAID layer from a reader process 810 of the extent store layer. Write operations are accumulated at a writer process 850 of the extent store layer, where associated extents 610 may be packed to form a full stripe before being dispatched to the RAID layer. The weight schedulers may be employed at the extent store layer to regulate the read operations and full stripe write operations at the RAID layer. Illustratively, a write weight scheduler 860 is provided for all SSDs 260 of a segment 650 at the writer process 850, whereas a read weight scheduler 820a-n is provided for every disk (e.g., SSD) of the segment 650 at the reader process 810 in order to control bandwidth distribution between the various categories of I/O operations.

Illustratively, there are a number of segments in the storage array that, once full, require segment cleaning (i.e., the number of clean segments have fallen below a free space threshold). The rate matching technique strives to segment clean at substantially the same rate as the incoming workload fills (i.e., consumes free space of) the segments. Otherwise, segment cleaning may proceed (1) too far ahead, thereby increasing write amplification or (2) fall behind. As for the latter, there may not be enough clean segments “in reserve”, thereby producing insufficient free space in segments to accept the incoming workload; in addition, a segment being cleaned may not be cleaned in time to accept the incoming workload. In the former case (i.e., cleaning too fast), finishing segment cleaning may yield lower latency to the host (i.e., negative “spike”). In both situations, a “spike” (negative or positive) in host-perceived latency may occur as a result of abrupt cleaning to free space (or completing cleaning) for the incoming workload. In an embodiment, each read weight scheduler 820a-n and write weight scheduler 860 may include a plurality of queues, wherein each queue is associated with one of the categories of I/O operations, e.g., user, metadata and relocation. In other words, the queues may include (i) a data queue for incoming user data (UD) workload (read 830a-n, write 870), (ii) a metadata queue (read 832a-n, write 872a-n) for incoming volume and extent store layer metadata (MD) workload, and (iii) a relocation queue (read 834a-n, write 874) for data and metadata relocation (REL) during segment cleaning. According to the rate matching technique, the weight scheduler may be configured to exert fine-grain control over the rate of segment cleaning to match the rate of the incoming I/O workload by assigning relative weights to the queues to thereby ensure that the segment cleaning does not introduce unnecessary write amplification (and increase host latency unnecessarily) by operating too quickly and does not impact performance that may increase a host-perceived latency associated with the I/O workload by operating too slowly. Note that a queue of each category (UD, MD, and REL) is provided for each SSD to support finer grain control of I/O.

To that end, it is desirable that each stripe written to the storage array (e.g., as part of a full stripe write operation 458) have a constant ratio of metadata, user data and relocation data, at steady-state; otherwise, the host-perceived latency may spike (i.e., exceed the bounded latency). The rate matching technique strives to provide smooth write latency, as well as read latency (e.g., with respect to user data read operations, metadata read operations and relocation read operations), i.e., to maintain smooth latency (without jitter-like behavior) to the host. Therefore, the technique balances read and write operations among user data, metadata and relocation data (I/O rate matching), as well as balancing (fine-grain control) read and write operations.

In an embodiment, the rate matching technique provides separate read queues 830a-n, 832a-n, 834a-n and write queues 870, 872, 874 that are scheduled by selecting and servicing a queue, i.e., processing operations (embodied as messages) from the queue. Notably, separate assemblages of read queues (UD, MD, and REL) associated with each respective read weight scheduler 820a-n may be assigned to each SSD; however, a single assemblage of write queues (UD, MD, and REL) may be assigned to the write weight scheduler 860. Illustratively, each SSD is assigned round-robin to the assemblage of write queues. A queue may be serviced (i.e., weight assigned) based on credit (amount of bytes serviced) that the queue has accumulated over time. For example, assume the weights of the three queues (user data, metadata, and relocation data) are assigned according to a ratio 1:2:3. Over a number “N” iterations, the weight scheduler strives to service a first queue (1/6*N), a second queue (2/6*N) and a third (3/6*N). Since the different categories of I/O operations may have different payload sizes (variable-sized extents), a “credit size” (e.g., 64 KB) worth of operations may be processed from the queue, as determined the read and write weight schedulers in order to maintain the queue weight ratio. In other words, the queues may be processed such that a number of bytes of I/O from each queue, rather than a number of queue entries, maintain the queue weight ratio.

Illustratively, there is one write weight scheduler 860 configured for write operations (i.e., shared among all SSDs), because the RAID layer performs full stripe write operations 458 across all SDDs. However, there may be a read weight scheduler 820a-n configured for read operations for each SSD, because extents are read from one SSD at a time and may be scattered differently across the SSDs. Thus, there is one read weight scheduler 820a-n per SSD 260a-n on the read path (based on single SSD read operations) that is implemented within the feedback loop on the SSD for the read queues, and there is one write weight scheduler 860 for all SSDs on the write path (based on full stripe write operations) that is implemented within the feedback loop across all the SSDs for the write queues. As noted, sources of the read operations (within the read path of the storage I/O stack) are illustratively: (i) user data (UD) read operations, (ii) metadata (MD) read operations from the volume and extent store layers and (iii) relocation (REL) read operations from the segment cleaning process. Similarly, these sources also exist for the write path of the storage I/O stack. With this framework, the I/O bandwidth of a certain class may be controlled by varying the weight of a corresponding queue to speed up or slow down segment relocation.

FIG. 9 illustrates the rate matching technique. In response to selecting a segment for cleaning, the segment cleaning process starts cleaning with a certain number of bytes to relocate, “bytes_to_relocate_at_start”. That is, as segment cleaning 910 is started, the segment has a total number of bytes in valid extents (i.e., not deleted or overwritten) to be relocated. Accounting structures 920, 940, e.g., one or more counters, may keep track of statistics such as the number of valid extents and/or deleted extents in a segment as extents are added (relocated by cleaning) and/or deleted (ignored by cleaning). These statistics may be used to derive a number of bytes that have been relocated by the segment cleaning process. Note that extents added to the segment increase the cleaning burden (i.e., increase the number of bytes to relocate); whereas, extents deleted from the segment decrease the cleaning burden (i.e., decrease the number of bytes to relocate). Progress of segment cleaning may be tracked in a segment cleaning progress accounting structure 920 having parameters (e.g., counters), such as “bytes_to_relocate_at_start” 922 and a number of “bytes_relocated” 924 to compute a percentage of bytes that have been relocated, “reloc_pct” 926, i.e., segment cleaning (relocation) progress:reloc_pct=(bytes_relocated/bytes_to_relocate_at_start)*100

Similarly, incoming I/O rate 930 (i.e., the incoming workload) for one or more segments selected for cleaning may be determined using free space availability. The number of bytes in a segment that do not have to be relocated is considered as free space, i.e., a quota of write space available (“write_quota_total”) for that segment. A number of bytes (i.e., running counter) written to/consumed for each segment since the segment cleaning process started may be tracked as “write_quota_used.” Accordingly, free space may be tracked using a free space accounting structure 940 having the parameters “write_quota_used” 944, “write_quota_total” 942, and a percentage of free space currently used (write_quota_pct) 946 computed as:write_quota_pct=(write_quota_used/write_quota_total)*100

Illustratively, as progress is recorded for cleaning (i.e., changes to percentage of bytes relocated) of a segment, progress is also recorded as to the user data and metadata written (i.e., changes to write space available), so that the rate of segment cleaning can be adjusted. That is, the amount of extents relocated from the segment being cleaned is tracked, as well as the amount of free space consumed in that segment. With these two parameters (i.e., reloc_pct 926 and write_quota_pct 946), a decision may be rendered as to whether to decrease or increase the rate of relocation. In other words, the parameters reloc_pct and write_quota_pct, may be used to feedback the incoming I/O rate and the segment cleaning rate. For example, assume 50% of extents need to be relocated when cleaning of a segment is commenced. Assume further during segment cleaning that incoming I/O may overwrite logical blocks, such that the amount of extents (blocks) to be relocated decreases, thereby allowing an increase in the incoming I/O rate (e.g., rate of incoming write requests from the host). The rate matching technique determines when and how to adjust the various input rates (e.g., to maintain smoothness and imperceptible fluctuations in latency to the host).

With information about the progress of segment cleaning (i.e., reloc_pct parameter 926) and available free space (i.e., write_quota_pct parameter 946), an expected weight of the relocation queue may be computed. The desired control (i.e., substantially matching the incoming I/O rate to the segment cleaning rate) may be achieved by computing an error between the rates, e.g., write_quota_pct_−reloc_pct, in an I/O rate matching controller 950 and applying that error to the expected weight of the relocation queues (e.g., write relocation queue 874) as an actuation to reduce the error, i.e., a feedback loop. Notably, the available free space acts as a negative feedback to limit the rate of segment cleaning, i.e., the more free space (less incoming I/O), the less cleaning is needed. Illustratively, the expected relocation weight is relative to a sum of the weights of the user and metadata queues. Once the expected weight is computed, the wght of the relocation queue may be adjusted in small increments or decrements, rather than directly instituting the expected weight, to avoid wide oscillations (e.g., control loop overshoot ringing) in weight that may cause jitter in performance (i.e., rapid variations in host-perceived latency). Note it is desirable to bound the latency (e.g., 1 millisecond) perceived by the host during control of the error, i.e., imbalance between the incoming I/O rate and relocation rate, so as to smooth latency variations. As will be understood by persons of skill in the art, many control algorithms, such as proportional-integrate-derivative (PID) control, may be applied to achieve the desired control.

The parameters that drive the segment cleaning rate (i.e., bytes_relocated 924 and write_quota_used 944) may be tracked and updated in the I/O path, whereas weight computation logic may be triggered when the difference in the relocation progress (reloc_pct 926) or free space usage (write_quota_pct 946) is more than 1%. That is, when either the relocation percentage (reloc_pct) or write quota percentage (write_quota_pct) changes by 1%, the weight (i.e., rate) computation of the I/O rate matching controller 950 is triggered. Accordingly, the control loop computation may be driven by a minimum change in the error rather than a fixed sampling rate. Separate processes (not shown) may update the parameters (i.e., the parameters of segment cleaning progress structure 920 and free space usage structure 940) as they change during control of the rates of the queues. Notably, the parameters may change as a result of segment cleaning (i.e., change in number of bytes relocated and amount free space) and/or incoming I/O operations (i.e., change in amount free space).

In an embodiment, a set of default weights may be assigned to the three queues, wherein a (fixed) weight of 50% may be assigned to the user data (UD) queue, a weight of 10% may be initially assigned to the metadata (MD) queue and a weight of 50% may be initially assigned to the relocation (REL) queue. Note the fixed weight assignment of the UD queue may be used to ensure at least a minimal servicing of host requests. The weight of the relocation queue may vary between, e.g., 10%-1000% (20 times greater than the UD queue), as the weights are typically driven by the fullness (amount of valid extents) of the segment, i.e., the fuller the segment, the greater the amount of extents that need to be relocated and, hence, the longer the relocation queue. For example, if the segment to be cleaned is 50% full, an amount of valid extents to be relocated (cleaned) should be similar to the amount of user data (and metadata) to be written in the time frame of filling a segment. Thus, weighting is dependent on the fullness of the storage array (segments), i.e., the amount of relocation needed to clean the segment is directly related to the fullness of the storage array. Note also that the rate of incoming I/O operations may depend upon on the amount of concurrency desired.

As used herein, concurrency is based on the incoming I/O workload, whereas rate matching is based on fullness of the storage array and progress (i.e., how far behind or ahead) of the relocation queues during segment cleaning. In particular, read concurrency may be determined based on RAID write requirements. In an embodiment, the rate matching technique may use the weight scheduler logic of the write process for concurrency of both read and write operations even though a separate weight scheduler of the read process is maintained for each disk (SSD), i.e., the weight of each read queue is identical. The processing of each queue is determined by the ratio of the user data (UD), metadata (MD) and relocation (REL) queues. If the storage array is full and there is a high rate of incoming user data I/O, segment cleaning is more heavily weighted to stay ahead of the I/O rate; otherwise, there would not be sufficient storage space to accommodate the incoming I/O workload. That is, if there is a high (undesirable) read latency during segment cleaning, relocation is not making as much progress as necessary to keep up with the incoming I/O write workload. An increase in the weight of the relocation queue may therefore be necessary so that the relocation queue may eventually “catch-up” to the rate of the incoming I/O workload.

Therefore, the rate matching technique strives to keep segment cleaning ahead of the I/O rate; otherwise the storage array may prematurely run out of (exhaust) the storage space. To that end, the relocation queue may be prioritized (through adjustments to the weights) with respect to the user data I/O workload. In addition, if user I/O operations are not received and processed by the storage I/O stack as fast as anticipated, the rate of segment cleaning may be reduced (slowed down) to avoid relocating (i.e., copying) extents that may not require such copying if cleaning of the segment is delayed (until necessary). Assume that the relocation queue is randomly prioritized to keep ahead of the incoming user data. As stripes are filled with extents for writing on the disks, a disproportionately lower amount of user data extents (e.g., less) may be assembled or packed into the stripes as compared to an amount of relocation extents (e.g., more). This may impact draining of NVRAM at the persistence layer, i.e., the servicing of the data in the NVRAM may not occur fast enough, resulting in “back pressure” to the host and a spike in host-perceived latency. The rate matching technique described herein attempts to avoid this situation so as to smooth-out latency and avoid such spikes by maintaining segment cleaning just ahead of the incoming I/O workload rate.

As noted, rate matching of queues for write workloads applies equally to read workloads. Depending on the read workloads, each SSD may have a different ratio of weighting ascribed to its queues. Illustratively, the relocation read queue rate is controlled to match the user data and metadata read (and write) queue rates. This results in bandwidth matching per SSD such that a SSD may be allocated less read operations when busy and more read operations when not busy. It is desirable to ensure that the number of read operations outstanding for the SSD due to relocation is sufficient so that segment cleaning (relocation) can be performed at a desired rate, while not impacting the smoothness of read latency of the host. In other words, there is no need to issue more read operations for relocation than is necessary to satisfy write operations for relocation. As such, the rate matching technique may be configured to prevent segment cleaning from impacting read operations of data and metadata by the user and storage I/O stack. Therefore, the rate of the read relocation queue may be controlled to match the rate of the write relocation queue such that the rates of the user and metadata queues are allowed to proceed as needed.

Weight calculation as based on feedback control described herein ensures that the relocation queue has the appropriate weight to maintain a desired segment cleaning progress. The services (e.g., reader and writer processes) that drive the relocation I/O operations are secondary control points that provide a finer grain of rate control 960 to track the relocation progress in order to process I/O operations at the desired rate, i.e., the rate set by the (primary) I/O rate matching controller 950 that sets the desired rate. The services that relocate extents take into account the following to process relocation I/O operations: (i) current weights of the queues at the writer process (i.e., desired rate set by the I/O rate matching); (ii) number of outstanding write requests (in bytes) of the three types of I/O operations (user, metadata, relocation) at the writer process; (iii) number of outstanding read requests (in bytes) across all of the per SSD weight schedulers at the reader process, and (iv) the relocation progress and/or free space availability. In addition, there are per-service limits to ensure that a service does not end up flooding I/O operations, wherein these limits include (v) number of relocation I/O operations dispatched per segment to be cleaned; (vi) maximum number of outstanding relocation I/O operations; and (vii) maximum number of outstanding relocation bytes.

While there have been shown and described illustrative embodiments directed to a rate matching technique configured to adjust a rate of cleaning of one or more selected segments of a storage array to accommodate a variable rate of incoming workload processed by the storage I/O stack executing on one or more nodes of a cluster, it is to be understood that various other adaptations and modifications may be made within the spirit and scope of the embodiments herein. For example, embodiments have been shown and described herein with relation to the rate matching technique employed at the extent store layer of the storage I/O stack. However, the embodiments in their broader sense are not so limited, and may, in fact, also allow for use of the rate matching technique in other storage applications, as well as at other layers of the storage I/O stack, such as the volume layer, where the technique may be used to implement de-reference count pacing (i.e., regulating deletes within the storage system).

The foregoing description has been directed to specific embodiments. It will be apparent, however, that other variations and modifications may be made to the described embodiments, with the attainment of some or all of their advantages. For instance, it is expressly contemplated that the components and/or elements described herein can be implemented as software encoded on a tangible (non-transitory) computer-readable medium (e.g., disks and/or CDs) having program instructions executing on a computer, hardware, firmware, or a combination thereof. Accordingly this description is to be taken only by way of example and not to otherwise limit the scope of the embodiments herein. Therefore, it is the object of the appended claims to cover all such variations and modifications as come within the true spirit and scope of the embodiments herein.

Claims

1. A method comprising: receiving, at an incoming rate, a plurality of write requests directed towards one or more logical units (LUN), each write request having data and processed at a node of a cluster, the node having a memory and connected to a storage array of solid state drives (SSDs);storing the data of each write request as one or more user data extents in a first segment spanning a set of SSDs, the first segment having a log-structured layout;tracking a percentage of free space of the first segment;tracking a percentage of bytes relocated from the first segment to a second segment spanning the set of SSDs, the second segment having the log-structured layout;computing an error by subtracting the percentage of free space from the percentage of bytes relocated; andcontrolling a rate of cleaning the first segment by substantially matching the cleaning rate to the incoming rate based on the computed error.

2. The method of claim 1 wherein the computed error is computed in response to a change in the percentage of free space exceeding a threshold.

3. The method of claim 1 further comprising: processing a relocation queue in proportion to a first weight assigned to the relocation queue, the first weight determined using the computed error, wherein the bytes relocated from the first segment to the second segment are queued to the relocation queue.

4. The method of claim 3 wherein the processing of the relocation queue in proportion to the first weight occurs by an amount of bytes processed from the relocation queue.

5. The method of claim 3 wherein the processing of the relocation queue in proportion to the first weight occurs by an amount of bytes processed from the relocation queue.

6. The method of claim 3 wherein the data of each write request is queued to a user data queue separate from the relocation queue.

7. The method claim 6 further comprising: processing the user data queue in proportion to a second weight assigned to the user data queue such that a ratio of the first weight to the second weight is maintained.

8. The method of claim 3 wherein each SSD of the set of SSDs has a separate read relocation queue, and wherein the first weight is assigned to a write relocation write queue shared among the set of SSDs.

9. The method of claim 8 wherein the first weight is modified in response to a number of read requests in the read queues exceeding a threshold.

10. A method comprising: receiving, at an incoming rate, a plurality of write requests directed towards one or more logical units (LUN), each write request having data and processed at a node of a cluster, the node having a memory and connected to a storage array of solid state drives (SSDs);storing the data of each write request as one or more user data extents in a first segment spanning a set of SSDs, the first segment having a log-structured layout;tracking an amount free space of the first segment;tracking an amount bytes relocated from the first segment to a second segment spanning the set of SSDs;setting a desired rate of relocating bytes from the first segment to the second segment by computing an error using the amount free space and the amount of bytes relocated; andcontrolling a relocation queue to maintain the desired rate of cleaning, wherein relocate bytes are queued to the relocation queue.

11. A system comprising: a storage system having a memory connected to a processor via a bus;a storage array coupled to the storage system and having one or more solid state drives (SSDs);a storage I/O stack executing on the processor of the storage system, the storage I/O stack configured to: receive, at an incoming rate, a plurality of write requests directed towards one or more logical units (LUN), each write request having data;store the data of each write request as one or more user data extents in a first segment spanning a set of SSDs, the first segment having a log-structured layout;track a percentage of free space of the first segment;track a percentage of bytes relocated from the first segment to a second segment spanning the set of SSDs, the second segment having the log-structured layout; andcompute an error by subtracting the percentage of free space from the percentage of bytes relocated; andcontrol a rate of cleaning the first segment by substantially matching the cleaning rate to the incoming rate based on the computed error.

12. The system of claim 11 wherein the computed error is computed in response to a change in the percentage of free space exceeding a threshold.

13. The system of claim 11 wherein the storage I/O stack is further configured to: process a relocation queue in proportion to a first weight assigned to the relocation queue, the first weight determined using the computed error, wherein the bytes relocated from the first segment to the second segment are queued to the relocation queue.

14. The system of claim 13 wherein the data of each write request is queued to a user data queue separate from the relocation queue.

15. The system of claim 14 wherein the storage I/O stack is further configured to: process the user data queue in proportion to a second weight assigned to the user data queue such that a ratio of the first weight to the second weight is maintained.

16. The system of claim 13 wherein each SSD of the set of SSDs has a separate read relocation queue, and wherein the first weight is assigned to a write queue shared among the set of SSDs.

17. The system of claim 16 wherein the first weight is modified in response to a number of read requests in the read queues exceeds a threshold.

18. The system of claim 16 wherein the first weight assigned to each write queue is modified in response to a number of relocation operations dispatched.