The present application is also related to U.S. patent application Ser. No. 11/119,278, entitled STORAGE SYSTEM ARCHITECTURE FOR STRIPING DATA CONTAINER CONTENT ACROSS VOLUMES OF A CLUSTER, by Michael Kazar, et al., the contents of which are hereby incorporated by reference.
The present invention relates to clustered computer environments and, more particularly, to guaranteeing data consistency to clients of a storage system cluster.
A storage system typically comprises one or more storage devices into which information may be entered, and from which information may be obtained, as desired. The storage system includes a storage operating system that functionally organizes the system by, inter alia, invoking storage operations in support of a storage service implemented by the system. The storage system may be implemented in accordance with a variety of storage architectures including, but not limited to, a network-attached storage environment, a storage area network and a disk assembly directly attached to a client or host computer. The storage devices are typically disk drives organized as a disk array, wherein the term “disk” commonly describes a self-contained rotating magnetic media storage device. The term disk in this context is synonymous with hard disk drive (HDD) or direct access storage device (DASD).
The storage operating system of the storage system may implement a high-level module, such as a file system, to logically organize the information stored on volumes as a hierarchical structure of data containers, such as files and logical units. For example, each “on-disk” file may be implemented as set of data structures, i.e., disk blocks, configured to store information, such as the actual data for the file. These data blocks are organized within a volume block number (vbn) space that is maintained by the file system. The file system may also assign each data block in the file a corresponding “file offset” or file block number (fbn). The file system typically assigns sequences of fbns on a per-file basis, whereas vbns are assigned over a larger volume address space. The file system organizes the data blocks within the vbn space as a “logical volume”; each logical volume may be, although is not necessarily, associated with its own file system.
A known type of file system is a write-anywhere file system that does not overwrite data on disks. If a data block is retrieved (read) from disk into a memory of the storage system and “dirtied” (i.e., updated or modified) with new data, the data block is thereafter stored (written) to a new location on disk to optimize write performance. A write-anywhere file system may initially assume an optimal layout such that the data is substantially contiguously arranged on disks. The optimal disk layout results in efficient access operations, particularly for sequential read operations, directed to the disks. An example of a write-anywhere file system that is configured to operate on a storage system is the Write Anywhere File Layout (WAFL®) file system available from Network Appliance, Inc., Sunnyvale, Calif.
The storage system may be further configured to operate according to a client/server model of information delivery to thereby allow many clients to access data containers stored on the system. In this model, the client may comprise an application, such as a database application, executing on a computer that “connects” to the storage system over a computer network, such as a point-to-point link, shared local area network (LAN), wide area network (WAN), or virtual private network (VPN) implemented over a public network such as the Internet. Each client may request the services of the storage system by issuing file-based and block-based protocol messages (in the form of packets) to the system over the network.
A plurality of storage systems may be interconnected to provide a storage system environment configured to service many clients. Each storage system may be configured to service one or more volumes, wherein each volume stores one or more data containers. Yet often a large number of data access requests issued by the clients may be directed to a small number of data containers serviced by a particular storage system of the environment. A solution to such a problem is to distribute the volumes serviced by the particular storage system among all of the storage systems of the environment. This, in turn, distributes the data access requests, along with the processing resources needed to service such requests, among all of the storage systems, thereby reducing the individual processing load on each storage system. However, a noted disadvantage arises when only a single data container, such as a file, is heavily accessed by clients of the storage system environment. As a result, the storage system attempting to service the requests directed to that file may exceed its processing resources and become overburdened, with a concomitant degradation of speed and performance.
One technique for overcoming the disadvantages of having a single file that is heavily utilized is to stripe the file across a plurality of volumes configured as a striped volume set (SVS), where each volume, such as a data volume (DV), is serviced by a different storage system, thereby distributing the load for the single file among a plurality of storage systems. A technique for data container (such as a file) striping is described in the above-referenced U.S. patent application Ser. No. 11/119,278, entitled STORAGE SYSTEM ARCHITECTURE FOR STRIPING DATA CONTAINER CONTENT ACROSS VOLUMES OF A CLUSTER. According to the data container striping arrangement, each storage system of may service access requests (i.e., file operations) from clients directed to the same file. File operations, such as read and write operations, are forwarded directly to the storage systems that are responsible for their portions of the data for that file.
In addition to the file data, there are meta-data, such as timestamps and length, associated with the file. A timestamp is a file attribute that provides an indication of the last time the file was modified, i.e., the modification time (mtime) for the file. The mtime is typically consulted on every operation directed to the file and, in the case of a write operation, is changed. For example, in response to a read operation issued by a client, the storage system returns the data and the current mtime on the file, whereas in response to a write operation, the storage system returns an incremented mtime. Effectively, every successive write operation is accorded a greater mtime than the one before it.
Many client protocols, such as the Network File System (NFS) protocol, allow use of client-side “caching” of data retrieved from a storage system. In response to a read operation issued by a client for a file, the storage system returns the requested data along with the current mtime of the file. The client stores the information in a cache memory so that future read operations directed to that file data may be serviced locally at the client (from the cache) instead of remotely over the network. For client-side caching to operate properly, there must be guarantees that the data subsequently retrieved from the cache is consistent with the actual file system and not “stale”, i.e., that the file data has not changed since it was cached at the client. To that end, the NFS protocol enables periodic “pinging” (polling) of the state of the file by the client through requests for the current mtime of the file from the storage system If the mtime has not increased since the data was cached, the client-side cache is maintained “fresh” and the client continues to use the cached data. If the mtime has changed, then the client discards its cached data and reissues a read operation to the storage system for file data.
Note that, as used herein, file operations are “serializable” if they can be replayed in a reported order and the result is identical to the actual file system. File operations are “causally connected” if they affect the same meta-data or the same region of the same file. Some client protocols (like NFSv2) require “strong serialization semantics”; that is, mtimes must always increase for operations that complete with increasing wall-clock time, even if they are not casually connected. “Weak serialization semantics”, on the other hand, only require that mtimes always increase for operation that complete with increasing wall-clock time if the operations are causally connected.
Certain file system protocols, such as the Common Internet File System (CIFS) protocol, support weak serialization semantics because of the nature of soft locks, such as opportunistic locks (op-locks). An op-lock is an automatically revocable soft lock that allows a client to operate on a range of file data until such time as a server (e.g., the storage system) instructs the client to stop. That is, the client can cache the data and perform read and write operations on the cached data until the storage system instructs it to return that data to the system. The client can cache the results of write operations since it knows that no other access is allowed to that same region of the file as long as it has an op-lock on the region. As soon as a second client attempts a conflicting operation on that region of the file, the storage system blocks the conflicting operation and revokes the op-lock. In particular, the storage system instructs the client to return (“flush”) any write modifications to the system and then discard the entire content of its client-side cache. Once that happens, the storage system unblocks the second client and grants it an op-lock to the conflicting region.
NFSv2 and NFSv3 protocols do not utilize op-locks and, thus, do not employ the above revocation system. For these protocols, the storage system must rely on strong serialization semantics. Other protocols, such as the NFSv4 protocol, use a type of soft lock called delegations that allows the storage system to use weak serialization semantics. Because CIFS and NFSv4 clients rely on such a “rough” protocol for guaranteeing consistency of cached data, they are not concerned with mtimes associated with read and write operations. This, in turn, enables the storage system to service such operation requests with weak serialization semantics.
In the data container striping arrangement described above, there is one volume, i.e., the container attribute volume (CAV), which is responsible for all the timestamps of a particular file stored on the SVS. As a result, for each file operation, the DV accesses the CAV to determine the mtime for the file. In response, the CAV updates the mtime on disk and returns the updated mtime to the DV which, in turn, returns the mtime and any associated data to the client. This arrangement places a substantial load on the storage system serving the CAV with a concomitant decrease in system performance. Moreover, depending on the load of the SVS, the meta-data requests to/from the CAV may become a bottleneck that adversely impacts performance of the system by, e.g., causing certain storage systems to stall (wait) until their meta-data requests have been processed before servicing client data access requests.
The present invention overcomes the disadvantages of the prior art by providing a system and method for efficiently guaranteeing data consistency to clients for one or more data containers stored on a plurality of volumes configured as a striped volume set (SVS) and served by a plurality of nodes, e.g., storage systems, connected as a cluster. The SVS comprises one meta-data volume (MDV) configured to store a canonical copy of certain meta-data, including access control lists and directories, associated with all data containers stored on the SVS, and one or more data volumes (DVs) configured to store, at least, data content of those containers. In addition, for each data container stored on the SVS, one volume is designated a container attribute volume (CAV) and, as such, is configured to store a canonical copy of certain, rapidly-changing attribute meta-data, including timestamps and length, associated with that container.
Efficient data consistency guarantees of data containers stored on the SVS are generally provided by delegating to the DVs sufficient authority to autonomously service input/output (I/O) requests directed to the data containers using at least some of the rapidly-changing attribute meta-data, e.g., the timestamps, of the containers. Specifically, a disk element of a storage system serving a DV (hereinafter “DV”) is only allowed to service I/O requests, e.g., read and write operations, to a data container, such as a file, if it has a valid ticket book for the file. The DV illustratively requests and is granted the ticket book from a disk element of the storage system serving the CAV (hereinafter “CAV”) on a per-file basis.
In the illustrative embodiment, the ticket book is a data structure generated by the CAV and comprising an indication of current timestamps, such as the current modification time (mtime), on the file plus zero or more “tickets”, i.e., new mtime values, that the DV is allowed to “hand out” (return) to a client for each new write operation. The types of ticket books illustratively include (i) a read ticket book that contains the current mtime and no tickets and (ii) a write ticket book that contains the current mtime and tickets representing a range of mtimes. Write operations require a write ticket book, while read and prefetch operations require at least a read ticket book. The write ticket book (hereinafter “ticket book”) is illustratively a read ticket book with one or more tickets that can be used to change the timestamps in response to write operations.
According to one aspect of the present invention, the ticket book is employed to improve storage system performance for clients that do not require strong serialization semantics. That is, for clients using file system protocols that support weak serialization semantics, the DV may utilize the ticket book in a manner that obviates the need to guarantee that the mtimes, as perceived by the clients, always increase. For clients using file system protocols that require strong serialization semantics, the DV may only use the ticket book if it received file operations prior to requesting (and granting of) that ticket book. Otherwise, the ticket book must be revoked and a new ticket book must be requested (and granted) from the CAV. Use of the ticket book in connection with weak serialization semantics thus reduces the number of round trip exchanges needed between the DV and CAV to service file operations. In other words, because the DV does not have to stall operation requests waiting for the grant of a new ticket book, the number of round trip exchanges is reduced compared to the number of round trip exchanges needed for protocols that require strong serialization semantics.
Another aspect of the invention is directed to the use of the ticket book with a file extending operation, e.g., a write operation that spans end-of-file (EOF) and increases the length of a file. In response to servicing a write operation that attempts to extend the file, the DV advises the CAV as to the new length of the file and, in return, the CAV grants a new ticket book reflective of that new file length. In particular, the CAV invalidates all outstanding ticket books to all DVs, updates the length of the file and returns the new ticket book to the advising DV. Since no other valid ticket books are in use, each DV must poll the CAV for an updated ticket book before servicing a new I/O operation. The new ticket book has tickets reflective of higher mtimes, and includes the correct new file length.
According to yet another aspect of the invention, a kinetic token is provided that represents an optimization that enables the storage system to defeat caching behavior at the client and improve the performance of file operations. A kinetic token is a guarantee that every time a client requests the current mtime on the file, the client will receive a higher value than has been previously reported. Issuance of a kinetic token by the CAV effectively disables client-side caching because every time the client requests the current mtime, it will receive a higher mtime value. Having disabled client-side caching in this manner, some protocols (such as NFSv3) can be satisfied with weak serialization semantics.
The above and further advantages of invention may be better understood by referring to the following description in conjunction with the accompanying drawings in which like reference numerals indicate identical or functionally similar elements:
A. Cluster Environment
The clients 180 may be general-purpose computers configured to interact with the node 200 in accordance with a client/server model of information delivery. That is, each client may request the services of the node, and the node may return the results of the services requested by the client, by exchanging packets over the network 140. The client may issue packets including file-based access protocols, such as the Common Internet File System (CIFS) protocol or Network File System (NFS) protocol, over the Transmission Control Protocol/Internet Protocol (TCP/IP) when accessing information in the form of files and directories. Alternatively, the client may issue packets including block-based access protocols, such as the Small Computer Systems Interface (SCSI) protocol encapsulated over TCP (iSCSI) and SCSI encapsulated over Fibre Channel (FCP), when accessing information in the form of blocks.
B. Storage System Node
Each node 200 is illustratively embodied as a dual processor storage system executing a storage operating system 300 that preferably implements a high-level module, such as a file system, to logically organize the information as a hierarchical structure of named data containers, such as directories, files and special types of files called virtual disks (hereinafter generally “blocks”) on the disks. However, it will be apparent to those of ordinary skill in the art that the node 200 may alternatively comprise a single or more than two processor system. Illustratively, one processor 222a executes the functions of the N-blade 310 on the node, while the other processor 222b executes the functions of the D-blade 350.
The memory 224 illustratively comprises storage locations that are addressable by the processors and adapters for storing software program code and data structures associated with the present invention. The processor and adapters may, in turn, comprise processing elements and/or logic circuitry configured to execute the software code and manipulate the data structures. The storage operating system 300, portions of which is typically resident in memory and executed by the processing elements, functionally organizes the node 200 by, inter alia, invoking storage operations in support of the storage service implemented by the node. It will be apparent to those skilled in the art that other processing and memory means, including various computer readable media, may be used for storing and executing program instructions pertaining to the invention described herein.
The network adapter 225 comprises a plurality of ports adapted to couple the node 200 to one or more clients 180 over point-to-point links, wide area networks, virtual private networks implemented over a public network (Internet) or a shared local area network. The network adapter 225 thus may comprise the mechanical, electrical and signaling circuitry needed to connect the node to the network. Illustratively, the computer network 140 may be embodied as an Ethernet network or a Fibre Channel (FC) network. Each client 180 may communicate with the node over network 140 by exchanging discrete frames or packets of data according to pre-defined protocols, such as TCP/IP.
The storage adapter 228 cooperates with the storage operating system 300 executing on the node 200 to access information requested by the clients. The information may be stored on any type of attached array of writable storage device media such as video tape, optical, DVD, magnetic tape, bubble memory, electronic random access memory, micro-electro mechanical and any other similar media adapted to store information, including data and parity information. However, as illustratively described herein, the information is preferably stored on the disks 130 of array 120. The storage adapter comprises a plurality of ports having input/output (I/O) interface circuitry that couples to the disks over an I/O interconnect arrangement, such as a conventional high-performance, FC link topology.
Storage of information on each array 120 is preferably implemented as one or more storage “volumes” that comprise a collection of physical storage disks 130 cooperating to define an overall logical arrangement of volume block number (vbn) space on the volume(s). Each logical volume is generally, although not necessarily, associated with its own file system. The disks within a logical volume/file system are typically organized as one or more groups, wherein each group may be operated as a Redundant Array of Independent (or Inexpensive) Disks (RAID). Most RAID implementations, such as a RAID-4 level implementation, enhance the reliability/integrity of data storage through the redundant writing of data “stripes” across a given number of physical disks in the RAID group, and the appropriate storing of parity information with respect to the striped data. An illustrative example of a RAID implementation is a RAID-4 level implementation, although it should be understood that other types and levels of RAID implementations may be used in accordance with the inventive principles described herein.
C. Storage Operating System
To facilitate access to the disks 130, the storage operating system 300 implements a write-anywhere file system that cooperates with one or more virtualization modules to “virtualize” the storage space provided by disks 130. The file system logically organizes the information as a hierarchical structure of named directories and files on the disks. Each “on-disk” file may be implemented as set of disk blocks configured to store information, such as data, whereas the directory may be implemented as a specially formatted file in which names and links to other files and directories are stored. The virtualization module(s) allow the file system to further logically organize information as a hierarchical structure of blocks on the disks that are exported as named logical unit numbers (luns).
In the illustrative embodiment, the storage operating system is preferably the NetApp® Data ONTAP™ operating system available from Network Appliance, Inc., Sunnyvale, Calif. that implements a Write Anywhere File Layout (WAFL™) file system. However, it is expressly contemplated that any appropriate storage operating system may be enhanced for use in accordance with the inventive principles described herein. As such, where the term “WAFL” is employed, it should be taken broadly to refer to any storage operating system that is otherwise adaptable to the teachings of this invention.
In addition, the storage operating system includes a series of software layers organized to form a storage server 365 that provides data paths for accessing information stored on the disks 130 of the node 200. To that end, the storage server 365 includes a file system module 360 in cooperating relation with a volume striping module (VSM) 370, a RAID system module 380 and a disk driver system module 390. The RAID system 380 manages the storage and retrieval of information to and from the volumes/disks in accordance with I/O operations, while the disk driver system 390 implements a disk access protocol such as, e.g., the SCSI protocol. The VSM 370 illustratively implements a striped volume set (SVS) described herein. As described further herein, the VSM co-operates with the file system 360 to enable storage server 365 to service a volume of the SVS. In particular, the VSM 370 implements a Locate( ) function 375 to compute the location of data container content in the SVS volume to thereby ensure consistency of such content served by the cluster.
The file system 360 implements a virtualization system of the storage operating system 300 through the interaction with one or more virtualization modules illustratively embodied as, e.g., a virtual disk (vdisk) module (not shown) and a SCSI target module 335. The vdisk module enables access by administrative interfaces, such as a user interface of a management framework 1110 (see
The file system 360 is illustratively a message-based system that provides logical volume management capabilities for use in access to the information stored on the storage devices, such as disks. That is, in addition to providing file system semantics, the file system 360 provides functions normally associated with a volume manager. These functions include (i) aggregation of the disks, (ii) aggregation of storage bandwidth of the disks, and (iii) reliability guarantees, such as mirroring and/or parity (RAID). The file system 360 illustratively implements the WAFL file system (hereinafter generally the “write-anywhere file system”) having an on-disk format representation that is block-based using, e.g., 4 kilobyte (kB) blocks and using index nodes (“inodes”) to identify files and file attributes (such as timestamps, access permissions, size and block location). The file system uses files to store meta-data describing the layout of its file system; these meta-data files include, among others, an inode file. A file handle, i.e., an identifier that includes an inode number, is used to retrieve an inode from disk.
Broadly stated, all inodes of the write-anywhere file system are organized into the inode file. A file system (fs) info block specifies the layout of information in the file system and includes an inode of a file that includes all other inodes of the file system. Each logical volume (file system) has an fsinfo block that is preferably stored at a fixed location within, e.g., a RAID group. The inode of the inode file may directly reference (point to) data blocks of the inode file or may reference indirect blocks of the inode file that, in turn, reference data blocks of the inode file. Within each data block of the inode file are embedded inodes, each of which may reference indirect blocks that, in turn, reference data blocks of a file.
Operationally, a request from the client 180 is forwarded as a packet over the computer network 140 and onto the node 200 where it is received at the network adapter 225. A network driver (of layer 312 or layer 330) processes the packet and, if appropriate, passes it on to a network protocol and file access layer for additional processing prior to forwarding to the write-anywhere file system 360. Here, the file system generates operations to load (retrieve) the requested data from disk 130 if it is not resident “in core”, i.e., in memory 224. If the information is not in memory, the file system 360 indexes into the inode file using the inode number to access an appropriate entry and retrieve a logical vbn. The file system then passes a message structure including the logical vbn to the RAID system 380; the logical vbn is mapped to a disk identifier and disk block number (disk,dbn) and sent to an appropriate driver (e.g., SCSI) of the disk driver system 390. The disk driver accesses the dbn from the specified disk 130 and loads the requested data block(s) in memory for processing by the node. Upon completion of the request, the node (and operating system) returns a reply to the client 180 over the network 140.
It should be noted that the software “path” through the storage operating system layers described above needed to perform data storage access for the client request received at the node may alternatively be implemented in hardware. That is, in an alternate embodiment of the invention, a storage access request data path may be implemented as logic circuitry embodied within a field programmable gate array (FPGA) or an application specific integrated circuit (ASIC). This type of hardware implementation increases the performance of the storage service provided by node 200 in response to a request issued by client 180. Moreover, in another alternate embodiment of the invention, the processing elements of adapters 225, 228 may be configured to offload some or all of the packet processing and storage access operations, respectively, from processor 222, to thereby increase the performance of the storage service provided by the node. It is expressly contemplated that the various processes, architectures and procedures described herein can be implemented in hardware, firmware or software.
As used herein, the term “storage operating system” generally refers to the computer-executable code operable on a computer to perform a storage function that manages data access and may, in the case of a node 200, implement data access semantics of a general purpose operating system. The storage operating system can also be implemented as a microkernel, an application program operating over a general-purpose operating system, such as UNIX® or Windows NT®, or as a general-purpose operating system with configurable functionality, which is configured for storage applications as described herein.
In addition, it will be understood to those skilled in the art that the invention described herein may apply to any type of special-purpose (e.g., file server, filer or storage serving appliance) or general-purpose node or computer, including a standalone computer or portion thereof, embodied as or including a storage system. Moreover, the teachings of this invention can be adapted to a variety of storage system architectures including, but not limited to, a network-attached storage environment, a storage area network and disk assembly directly-attached to a client or host computer. The term “storage system” should therefore be taken broadly to include such arrangements in addition to any subsystems configured to perform a storage function and associated with other equipment or systems. It should be noted that while this description is written in terms of a write any where file system, the teachings of the present invention may be utilized with any suitable file system, including a write in place file system.
D. CF Protocol
In the illustrative embodiment, the storage server 365 is embodied as D-blade 350 of the storage operating system 300 to service one or more volumes of array 120. In addition, the multi-protocol engine 325 is embodied as N-blade 310 to (i) perform protocol termination with respect to a client issuing incoming data access request packets over the network 140, as well as (ii) redirect those data access requests to any storage server 365 of the cluster 100. Moreover, the N-blade 310 and D-blade 350 cooperate to provide a highly-scalable, distributed storage system architecture of the cluster 100. To that end, each blade includes a cluster fabric (CF) interface module 340a,b adapted to implement intra-cluster communication among the blades, including D-blade-to-D-blade communication, for data container striping operations described herein.
The protocol layers, e.g., the NFS/CIFS layers and the iSCSI/FC layers, of the N-blade 310 function as protocol servers that translate file-based and block based data access requests from clients into CF protocol messages used for communication with the D-blade 350. That is, the N-blade servers convert the incoming data access requests into file system primitive operations (commands) that are embedded within CF messages by the CF interface module 340 for transmission to the D-blades 350 of the cluster 100. Notably, the CF interface modules 340 cooperate to provide a single file system image across all D-blades 350 in the cluster 100. Thus, any network port of an N-blade that receives a client request can access any data container within the single file system image located on any D-blade 350 of the cluster.
Further to the illustrative embodiment, the N-blade 310 and D-blade 350 are implemented as separately-scheduled processes of storage operating system 300; however, in an alternate embodiment, the blades may be implemented as pieces of code within a single operating system process. Communication between an N-blade and D-blade is thus illustratively effected through the use of message passing between the blades although, in the case of remote communication between an N-blade and D-blade of different nodes, such message passing occurs over the cluster switching fabric 150. A known message-passing mechanism provided by the storage operating system to transfer information between blades (processes) is the Inter Process Communication (IPC) mechanism. The protocol used with the IPC mechanism is illustratively a generic file and/or block-based “agnostic” CF protocol that comprises a collection of methods/functions constituting a CF application programming interface (API). Examples of such an agnostic protocol are the SpinFS and SpinNP protocols available from Network Appliance, Inc. The SpinFS protocol is described in the above-referenced U.S. Patent Application Publication No. US 2002/0116593.
The CF interface module 340 implements the CF protocol for communicating file system commands among the blades of cluster 100. Communication is illustratively effected by the D-blade exposing the CF API to which an N-blade (or another D-blade) issues calls. To that end, the CF interface module 340 is organized as a CF encoder and CF decoder. The CF encoder of, e.g., CF interface 340a on N-blade 310 encapsulates a CF message as (i) a local procedure call (LPC) when communicating a file system command to a D-blade 350 residing on the same node 200 or (ii) a remote procedure call (RPC) when communicating the command to a D-blade residing on a remote node of the cluster 100. In either case, the CF decoder of CF interface 340b on D-blade 350 de-encapsulates the CF message and processes the file system command.
A data container, e.g., a file, is accessed in the file system using a data container handle.
E. File System Organization
In the illustrative embodiment, a data container is represented in the write-anywhere file system as an inode data structure adapted for storage on the disks 130.
Specifically, the data section 660 of a regular on-disk inode may include file system data or pointers, the latter referencing 4 kB data blocks on disk used to store the file system data. Each pointer is preferably a logical vbn to facilitate efficiency among the file system and the RAID system 380 when accessing the data on disks. Given the restricted size (e.g., 128 bytes) of the inode, file system data having a size that is less than or equal to 64 bytes is represented, in its entirety, within the data section of that inode. However, if the length of the contents of the data container exceeds 64 bytes but less than or equal to 64 kB, then the data section of the inode (e.g., a first level inode) comprises up to 16 pointers, each of which references a 4 kB block of data on the disk.
Moreover, if the size of the data is greater than 64 kB but less than or equal to 64 megabytes (MB), then each pointer in the data section 660 of the inode (e.g., a second level inode) references an indirect block (e.g., a first level L1 block) that contains 1024 pointers, each of which references a 4 kB data block on disk. For file system data having a size greater than 64 MB, each pointer in the data section 660 of the inode (e.g., a third level L3 inode) references a double-indirect block (e.g., a second level L2 block) that contains 1024 pointers, each referencing an indirect (e.g., a first level L1) block. The indirect block, in turn, contains 1024 pointers, each of which references a 4 kB data block on disk. When accessing a file, each block of the file may be loaded from disk 130 into the memory 224.
When an on-disk inode (or block) is loaded from disk 130 into memory 224, its corresponding in-core structure embeds the on-disk structure. For example, the dotted line surrounding the inode 600 indicates the in-core representation of the on-disk inode structure. The in-core structure is a block of memory that stores the on-disk structure plus additional information needed to manage data in the memory (but not on disk). The additional information may include, e.g., a “dirty” bit 670. After data in the inode (or block) is updated/modified as instructed by, e.g., a write operation, the modified data is marked “dirty” using the dirty bit 670 so that the inode (block) can be subsequently “flushed” (stored) to disk. The in-core and on-disk format structures of the WAFL file system, including the inodes and inode file, are disclosed and described in the previously incorporated U.S. Pat. No. 5,818,292 titled METHOD FOR MAINTAINING CONSISTENT STATES OF A FILE SYSTEM AND FOR CREATING USER-ACCESSIBLE READ-ONLY COPIES OF A FILE SYSTEM by David Hitz et al.
A file system layout is provided that apportions an underlying physical volume into one or more virtual volumes (or flexible volumes) of a storage system, such as node 200. An example of such a file system layout is described in U.S. patent application Ser. No. 10/836,817 titled EXTENSION OF WRITE ANYWHERE FILE SYSTEM LAYOUT, by John K. Edwards et al. and assigned to Network Appliance, Inc. The underlying physical volume is an aggregate comprising one or more groups of disks, such as RAID groups, of the node. The aggregate has its own physical volume block number (pvbn) space and maintains meta-data, such as block allocation structures, within that pvbn space. Each flexible volume has its own virtual volume block number (vvbn) space and maintains meta-data, such as block allocation structures, within that vvbn space. Each flexible volume is a file system that is associated with a container file; the container file is a file in the aggregate that contains all blocks used by the flexible volume. Moreover, each flexible volume comprises data blocks and indirect blocks that contain block pointers that point at either other indirect blocks or data blocks.
In one embodiment, pvbns are used as block pointers within buffer trees of files (such as file 700) stored in a flexible volume. This “hybrid” flexible volume embodiment involves the insertion of only the pvbn in the parent indirect block (e.g., inode or indirect block). On a read path of a logical volume, a “logical” volume (vol) info block has one or more pointers that reference one or more fsinfo blocks, each of which, in turn, points to an inode file and its corresponding inode buffer tree. The read path on a flexible volume is generally the same, following pvbns (instead of vvbns) to find appropriate locations of blocks; in this context, the read path (and corresponding read performance) of a flexible volume is substantially similar to that of a physical volume. Translation from pvbn-to-disk,dbn occurs at the file system/RAID system boundary of the storage operating system 300.
In an illustrative dual vbn hybrid flexible volume embodiment, both a pvbn and its corresponding vvbn are inserted in the parent indirect blocks in the buffer tree of a file. That is, the pvbn and vvbn are stored as a pair for each block pointer in most buffer tree structures that have pointers to other blocks, e.g., level 1 (L1) indirect blocks, inode file level 0 (L0) blocks.
The pvbns reference locations on disks of the aggregate, whereas the vvbns reference locations within files of the flexible volume. The use of pvbns as block pointers 808 in the indirect blocks 804 provides efficiencies in the read paths, while the use of vvbn block pointers provides efficient access to required meta-data. That is, when freeing a block of a file, the parent indirect block in the file contains readily available vvbn block pointers, which avoids the latency associated with accessing an owner map to perform pvbn-to-vvbn translations; yet, on the read path, the pvbn is available.
Whereas the aggregate 900 is analogous to a physical volume of a conventional storage system, a flexible volume is analogous to a file within that physical volume. That is, the aggregate 900 may include one or more files, wherein each file contains a flexible volume 910 and wherein the sum of the storage space consumed by the flexible volumes is physically smaller than (or equal to) the size of the overall physical volume. The aggregate utilizes a physical pvbn space that defines a storage space of blocks provided by the disks of the physical volume, while each embedded flexible volume (within a file) utilizes a logical vvbn space to organize those blocks, e.g., as files. Each vvbn space is an independent set of numbers that corresponds to locations within the file, which locations are then translated to dbns on disks. Since the flexible volume 910 is also a logical volume, it has its own block allocation structures (e.g., active, space and summary maps) in its vvbn space.
A container file is a file in the aggregate that contains all blocks used by a flexible volume. The container file is an internal (to the aggregate) feature that supports a flexible volume; illustratively, there is one container file per flexible volume. Similar to a pure logical volume in a file approach, the container file is a hidden file (not accessible to a user) in the aggregate that holds every block in use by the flexible volume. The aggregate includes an illustrative hidden meta-data root directory that contains subdirectories of flexible volumes:
Specifically, a physical file system (WAFL) directory includes a subdirectory for each flexible volume in the aggregate, with the name of subdirectory being a file system identifier (fsid) of the flexible volume. Each fsid subdirectory (flexible volume) contains at least two files, a filesystem file and a storage label file. The storage label file is illustratively a 4 kB file that contains meta-data similar to that stored in a conventional raid label. In other words, the storage label file is the analog of a raid label and, as such, contains information about the state of the flexible volume such as, e.g., the name of the flexible volume, a universal unique identifier (uuid) and fsid of the flexible volume, whether it is online, being created or being destroyed, etc.
In addition to being embodied as a container file having level 1 blocks organized as a container map, the filesystem file 1040 includes block pointers that reference various file systems embodied as flexible volumes 1050. The aggregate 1000 maintains these flexible volumes 1050 at special reserved inode numbers. Each flexible volume 1050 also has special reserved inode numbers within its flexible volume space that are used for, among other things, the block allocation bitmap structures. As noted, the block allocation bitmap structures, e.g., active map 1062, summary map 1064 and space map 1066, are located in each flexible volume.
Specifically, each flexible volume 1050 has the same inode file structure/content as the aggregate, with the exception that there is no owner map and no WAFL/fsid/filesystem file, storage label file directory structure in a hidden meta-data root directory 1080. To that end, each flexible volume 1050 has a volinfo block 1052 that points to one or more fsinfo blocks 1054, each of which may represent a snapshot along with the active file system of the flexible volume. Each fsinfo block, in turn, points to an inode file 1060 that, as noted, has the same inode structure/content as the aggregate with the exceptions noted above. Each flexible volume 1050 has its own inode file 1060 and distinct inode space with corresponding inode numbers, as well as its own root (fsid) directory 1070 and subdirectories of files that can be exported separately from other flexible volumes.
The storage label file 1090 contained within the hidden meta-data root directory 1030 of the aggregate is a small file that functions as an analog to a conventional raid label. A raid label includes physical information about the storage system, such as the volume name; that information is loaded into the storage label file 1090. Illustratively, the storage label file 1090 includes the name 1092 of the associated flexible volume 1050, the online/offline status 1094 of the flexible volume, and other identity and state information 1096 of the associated flexible volume (whether it is in the process of being created or destroyed).
F. VLDB
The VLDB 1130 is a database process that tracks the locations of various storage components (e.g., SVSs, flexible volumes, aggregates, etc.) within the cluster 100 to thereby facilitate routing of requests throughout the cluster. In the illustrative embodiment, the N-blade 310 of each node accesses a configuration table 235 that maps the SVS ID 502 of a data container handle 500 to a D-blade 350 that “owns” (services) the data container within the cluster. The VLDB includes a plurality of entries which, in turn, provide the contents of entries in the configuration table 235; among other things, these VLDB entries keep track of the locations of the flexible volumes (hereinafter generally “volumes 910”) and aggregates 900 within the cluster. Examples of such VLDB entries include a VLDB volume entry 1200 and a VLDB aggregate entry 1300.
The VLDB illustratively implements a RPC interface, e.g., a Sun RPC interface, which allows the N-blade 310 to query the VLDB 1130. When encountering contents of a data container handle 500 that are not stored in its configuration table, the N-blade sends an RPC to the VLDB process. In response, the VLDB 1130 returns to the N-blade the appropriate mapping information, including an ID of the D-blade that owns the data container. The N-blade caches the information in its configuration table 235 and uses the D-blade ID to forward the incoming request to the appropriate data container. All functions and interactions between the N-blade 310 and D-blade 350 are coordinated on a cluster-wide basis through the collection of management processes and the RDB library user mode applications 1100. When processing an SVS, the N-blade caches the striping rules that determine which stripes of data lie in which container.
To that end, the management processes have interfaces to (are closely coupled to) RDB 1150. The RDB comprises a library that provides a persistent object store (storing of objects) for the management data processed by the management processes. Notably, the RDB 1150 replicates and synchronizes the management data object store access across all nodes 200 of the cluster 100 to thereby ensure that the RDB database image is identical on all of the nodes 200. At system startup, each node 200 records the status/state of its interfaces and IP addresses (those IP addresses it “owns”) into the RDB database.
G. Storage System Architecture
The present invention is related to a storage system architecture illustratively comprising two or more volumes 910 distributed across a plurality of nodes 200 of cluster 100. The volumes are organized as a SVS and configured to store content of data containers, such as files and luns, served by the cluster in response to multi-protocol data access requests issued by clients 180. Notably, the content of each data container is apportioned among the volumes of the SVS to thereby improve the efficiency of storage service provided by the cluster. To facilitate a description and understanding of the present invention, data containers are hereinafter referred to generally as “files”.
The SVS comprises a meta-data volume (MDV) and one or more data volumes (DV). The MDV is configured to store a canonical copy of certain meta-data, including access control lists (ACLs) and directories, associated with all files stored on the SVS, whereas each DV is configured to store, at least, data content of those files. For each file stored on the SVS, one volume is designated the container attribute volume (CAV) and, to that end, is configured to store (“cache”) certain, rapidly-changing attribute meta-data, including time stamps and file length, associated with that file to thereby offload access requests that would otherwise be directed to the MDV.
In the illustrative embodiment described herein, determination of the CAV for a file is based on a simple rule: designate the volume holding the first stripe of content (data) for the file as the CAV for the file. Not only is this simple rule convenient, but it also provides an optimization for small files. That is, a CAV may be able to perform certain operations without having to communicate with other volumes of the SVS if the file is small enough to fit within the specified stripe width. Ideally, the first stripes of data for files are distributed among the DVs of the SVS to thereby facilitate even distribution of CAV designations among the volumes of the SVS. In alternate embodiments, data for files is striped across the MDV and the DVs.
In contrast, each DV 1410, 1415 stores only file (F) inodes 1425, 1435, 1445 and ACL inode 1440. According to the inventive architecture, a DV does not store directories or other device inodes/constructs, such as symbolic links; however, each DV does store F inodes, and may store cached copies of ACL inodes, that are arranged in the same locations as their respective inodes in the MDV 1405. A particular DV may not store a copy of an inode until an I/O request for the data container associated with the inode is received by the D-blade serving a particular DV. Moreover, the contents of the files denoted by these F inodes are periodically sparse according to SVS striping rules, as described further herein. In addition, since one volume is designated the CAV for each file stored on the SVS 1400, DV 1415 is designated the CAV for the file represented by inode 1425 and DV 1410 is the CAV for the files identified by inodes 1435, 1445. Accordingly, these CAVs cache certain, rapidly-changing attribute meta-data (M) associated with those files such as, e.g., file size 615, as well as access and/or modification time (mtime) stamps 620.
The SVS is associated with a set of striping rules that define a stripe algorithm, a stripe width and an ordered list of volumes within the SVS. The striping rules for each SVS are illustratively stored as an entry of VLDB 1130 and accessed by SVS ID.
Each set of striping rules 1530 illustratively includes a stripe width field 1510, a stripe algorithm ID field 1515, an ordered list of volumes field 1520 and, in alternate embodiments, additional fields 1525. The striping rules 1530 contain information for identifying the organization of a SVS. For example, the stripe algorithm ID field 1515 identifies a striping algorithm used with the SVS. In the illustrative embodiment, multiple striping algorithms could be used with a SVS; accordingly, stripe algorithm ID is needed to identify which particular algorithm is utilized. Each striping algorithm, in turn, specifies the manner in which file content is apportioned as stripes across the plurality of volumes of the SVS. The stripe width field 1510 specifies the size/width of each stripe. The ordered list of volumes field 1520 contains the IDs of the volumes comprising the SVS. Moreover, the ordered list of volumes may specify the function and implementation of the various volumes and striping rules of the SVS. For example, the first volume in the ordered list may denote the MDV of the SVS, whereas the ordering of volumes in the list may denote the manner of implementing a particular striping algorithm, e.g., round-robin.
A Locate( ) function 375 is provided that enables the VSM 370 and other modules (such as those of N-blade 310) to locate a D-blade 350 and its associated volume of a SVS 1400 in order to service an access request to a file. The Locate( ) function takes as arguments, at least (i) a SVS ID 1505, (ii) an offset within the file, (iii) the inode number for the file and (iv) a set of striping rules 1530, and returns the volume 910 on which that offset begins within the SVS 1400. For example, assume a data access request directed to a file is issued by a client 180 and received at the N-blade 310 of a node 200, where it is parsed through the multi-protocol engine 325 to the appropriate protocol server of N-blade 310. To determine the location of a D-blade 350 to which to transmit a CF message 400, the N-blade 310 may first retrieve a SVS entry 1500 to acquire the striping rules 1530 (and list of volumes 1520) associated with the SVS. The N-blade 310 then executes the Locate( ) function 375 to identify the appropriate volume to which to direct an operation. Thereafter, the N-Blade may retrieve the appropriate VLDB volume entry 1200 to identify the aggregate containing the volume and the appropriate VLDB aggregate entry 1300 to ultimately identify the appropriate D-blade 350. The protocol server of N-blade 310 then transmits the CF message 400 to the D-blade 350.
In accordance with an illustrative round robin striping algorithm, volume A 1605 contains a stripe of file content or data (D) 1620 followed, in sequence, by two stripes of sparseness (S) 1622, 1624, another stripe of data (D) 1626 and two stripes of sparseness (S) 1628, 1630. Volume B 1610, on the other hand, contains a stripe of sparseness (S) 1632 followed, in sequence, by a stripe of data (D) 1634, two stripes of sparseness (S) 1636, 1638, another stripe of data (D) 1640 and a stripe of sparseness (S) 1642. Volume C 1615 continues the round robin striping pattern and, to that end, contains two stripes of sparseness (S) 1644, 1646 followed, in sequence, by a stripe of data (D) 1648, two stripes of sparseness (S) 1650, 1652 and another stripe of data (D) 1654.
H. Data Consistency Guarantees
The present invention is directed to a system and method for efficiently guaranteeing data consistency to clients for one or more data containers stored on a plurality of volumes configured as a SVS. As noted, the SVS comprises one MDV configured to store a canonical copy of certain meta-data, including access control lists and directories, associated with all data containers stored on the SVS, and one or more DVs configured to store, at least, data content of those containers. In addition, for each data container stored on the SVS, one volume is designated the CAV and, as such, is configured to store certain, rapidly-changing attribute meta-data, including timestamps and length, associated with that container.
Twelve data containers, e.g., files (files 1-12), are illustratively stored on the volumes of the SVS, wherein each volume serves as the CAV for any file whose first stripe is stored therein. Notably, the CAV is a role that the MDV or DV serves for a particular file to store (and serve) rapidly-changing attribute meta-data for the file. Thus, for example, volume B 1710 serves as the CAV for files 1, 5, 9. Similarly, volume C 1715 serves as the CAV for files 2, 6 and 10, volume D 1720 serves as the CAV for files 3, 7 and 11 and volume E serves as the CAV for files 4, 8 and 12. Volume A 1705, which serves as the MDV for the SVS does not, in the illustrative embodiment, serve as the CAV for any files. In alternate embodiments, the MDV may serve as the CAV for files. By distributing the role of the CAV among the SVS volumes, each volume serves as a CAV for an approximately equal number of files.
The meta-data associated with the files stored on the SVS are illustratively organized into various categories (e.g., MD1-MD3) along functional boundaries and are resident on various volumes to optimize data access (e.g., read and write) paths through the SVS. These categories include (i) MD1 meta-data that changes on every data access (read/write) request served by the SVS, (ii) MD2 meta-data that may be retrieved (but not changed) on every request and (iii) MD3 meta-data that is unused for the read/write requests. Since it changes on every read/write request served by the DVs of the SVS, the MD1 meta-data is canonically resident on the CAV and generally cached on the DVs. Likewise, since it may be retrieved, but does not change, on every request served by the DVs, the MD2 meta-data is canonically resident on the MDV and generally cached on all DVs of the SVS, including the volume designated as CAV. Finally, since it is unused for a read/write request, the MD3 meta-data is canonically, and solely, resident on the MDV.
In the illustrative embodiment, the CAV cooperates with the MDV and DVs of the SVS to provide a multi-tier caching and distribution architecture that offloads meta-data access requests that would otherwise be directed to the MDV.
As noted above, every volume within the SVS serves as a CAV for a portion of the files stored on the SVS. Illustratively, the volume storing the first stripe of the file is deemed to be the CAV. Thus, different volumes of the environment 1800 may serve as the CAV 1810 depending on the particular files. Moreover, the MDV may serve as the CAV for certain files. In such a case, the roles of the CAV and MDV are merged due to the fact that the volume contains a canonical copy of all (MD1-MD3) meta-data. For those files stored on a SVS volume having a merged CAV/MDV role, no first tier caching and distribution point is utilized in the hierarchical environment 1800.
Data consistency guarantees of data containers, e.g., files, stored on the SVS is generally provided by delegating to the DVs sufficient authority to autonomously service I/O requests directed to the files using at least some of the rapidly-changing attribute meta-data, e.g., the timestamps, of the files. As noted, each node includes a D-blade 350 configured to service a volume, such as the MDV 1805, DV 1815 or CAV 1810. To further facilitate a description and understanding of the present invention, references to the “MDV”, “DV” and “CAV” may include the D-blades configured to service those volumes. Specifically, a DV 1815 is only allowed to service I/O requests, e.g., read and write operations, to a file if it has a valid ticket book for the file. A DV requests and is granted the ticket book from the CAV 1810 on a per-file basis, although a DV may have any number of allocated ticket books based on the number of outstanding or currently active files.
According to one aspect of the present invention, the ticket book 1900 is employed to improve storage system performance for clients that do not require strong serialization semantics. That is, for clients using file system protocols that support weak serialization semantics, such as CIFS, the DV 1815 may utilize the ticket book in a manner that obviates the need to guarantee that the mtimes, as perceived by the clients, always increase. For clients using file system protocols that require strong serialization semantics, the DV may only use the ticket book if it received file operations prior to requesting (and granting of) that ticket book. Otherwise, a new ticket book must be requested (and granted) from the CAV 1810 to process the received file operations. Use of the ticket book in connection with weak serialization semantics thus reduces the number of round trip exchanges needed between the DV and CAV to service file operations. In other words, because the DV does not have to stall operation requests waiting for the grant of a new ticket book, the number of round trip exchanges is reduced compared to the number of round trip exchanges needed for protocols that require strong serialization semantics.
When servicing a read operation, each DV returns the lowest mtime in its current ticket book to the requesting client. When servicing a write operation, however, the DV increases (e.g., increments) the mtime before returning it to the client. Moreover, the DV assigns a ticket 1920 from the ticket book 1900 to the write operation atomically with the step of committing the write to disk. These two aspects combine to guarantee that every new write operation that a DV completes will be assigned a post-operation mtime value that is higher than any mtime that the DV has associated with any earlier operation.
In addition, each DV allocates the mtimes of its delegated range in sequence. For example, assume DV1 has a ticket book for file A with an mtime range of 101 to 150. In response to a first read operation directed to the file, DV1 returns mtime 101 with the data. Subsequently, in response to a first write operation directed to the file, DV1 increments the mtime from 101 to 102 and returns mtime 102 with the request's post-operation attributes. DV1 then removes ticket 1920 reflecting mtime 102 from its book such that the new range is 103 to 150. Any subsequent read operations directed to the file return current mtimes of 102 until a second write operation is issued, at which time the incremented mtime 103 is returned. When it depletes those mtimes, i.e., when it returns mtime 150 and runs out of tickets, DV1 requests a new ticket book from the CAV. Alternatively, if a DV only has a read ticket book, then it must obtain a new range of the timeline from the CAV before it can service any more write operations.
Notably, there is no guarantee that the mtime always reflects “wall clock” time, i.e., the actual time of day. That is, in some circumstances it is possible (and acceptable) for a particular client to issue an I/O request (e.g., a first write operation) and receive an acknowledgment with a first mtime (e.g., 100) and then synchronously issue a second write operation and receive a smaller mtime (e.g., 50) even though the client has knowledge that the second write operation occurred after the first write operation. For example, assume that a client issues a first write operation to a first region (e.g., a stripe) of a file that is serviced by DV2 and waits for a response before proceeding with any further operations. DV2 performs the first write operation and returns mtime 201 to client for that operation. The client then issues a second write operation to a second, different stripe of the file that is serviced by DV1. DV1 performs the second write operation and returns mtime 101 to the client for that operation. Essentially, the client perceives the returned mtimes as proceeding backwards; the mtime for the second write operation is “earlier” even though the client is aware that the first write operation fully completed before the second operation began.
Certain client protocols, such as NFSv2, cannot support this situation and thus require “strong serialization semantics”. As used herein, all file operations are serializable using a timeline indexed by a file's mtime. For causally connected operations, increasing mtimes correlate with increasing wall clock time. If a client protocol requires strong serialization semantics, increasing mtimes correlate with increasing wall clock time. However, those client protocols that can support the situation described above accept “weak serialization semantics” and the present invention provides optimizations to accommodate those protocols. CIFS and, to some extent, NFSv4 client protocols are configured to support weak serialization semantics; accordingly, the ticket book 1900 provides an optimization in support of such semantics. That is, weak serialization semantics allow DV1 to exploit the use of its ticket book by, e.g., obviating the need to guarantee that the mtime, as perceived by the client, always increases. Note that this situation only manifests when a client's operation requests traverse DVs, i.e., the client transitions from writing to one DV to writing to another DV. As long as the client directs requests to one DV, the mtimes always increase.
In the illustrative embodiment, there are four basic rules governing when a ticket book 1900 currently held by a DV 1815 can be used. If any of these rules fails, then the DV has to obtain a new one from the CAV 1810 before servicing certain operation requests. According to a first rule, a DV must have sufficient “up-to-date” meta-data to service an I/O operation to a file. Otherwise if the meta-data is out-of-date for the file, the DV must obtain that meta-data and, in the process of obtaining that meta-data, the DV obtains a new ticket book. In general, whenever a DV 1815 accesses the CAV 1810 for additional attributes or meta-data, the CAV issues a new ticket book to the DV.
A second rule states that if the ticket book 1900 expires or is otherwise unusable (i.e., all tickets 1920 are exhausted) such that no further write operations can be performed, then the DV must obtain a new ticket book. Here, the ticket book itself becomes a kind of meta-data that the DV may need to retrieve from the CAV. A ticket book may expire based on wall clock time. It is preferable that the timestamps returned to a client be generally close to actual wall clock time. After a predetermined time (e.g., 100 milliseconds or approximately 10 times a second), the DV discards its current ticket book and obtains a new ticket book the next time a request is issued to the DV. Notably, the new ticket book is obtained “on demand”, i.e., when a request is received at the DV that cannot be satisfied using the current ticket book.
Illustratively, an optimization may be invoked for this second rule. If there may still be activity directed to the file and the current ticket book for the file is set to expire shortly, the DV can proactively request a new ticket book for the file so that by the time that the next request is received, the DV has the new ticket book. This optimization is based on a heuristic to decide how frequently the DV should proactively obtain a ticket book without being prompted by an incoming client request. This optimization is similar to a “meta-data” read ahead operation that prepares for a new ticket book in case one is needed. However, if proactive requests are performed too aggressively, the CAV would be overloaded.
A number of different algorithms may be employed that specify when a DV should approach the CAV for a new ticket book. An example of an illustrative algorithm specifies that if at least one I/O operation to a particular file is serviced using a current ticket book for that file and if the ticket book is about to expire within a predetermined time equal to the round trip time to the CAV (e.g., the next 20 milliseconds), then the DV proactively obtains a new ticket book from the CAV.
A third rule involves a situation where an I/O operation, such as a write operation, spans the end of file (EOF). Write operations that span the current EOF change the length of the file. File-length changes are considered causally connected with all other I/O operations; that is, if a file is extended by a write operation and that operation returns post-operation mtime X to the client, then all subsequent I/O operations return mtime values greater than X and all these subsequent operations use the correct, new file length. This is accomplished by providing the following third rule: when a DV 1815 wants to extend the file, it must always go to the CAV 1810 to request a new ticket book.
Therefore, this aspect of the invention is directed to the use of the ticket book with a file extending operation, e.g., a write operation that spans EOF and increases the length of a file. In response to servicing a write operation that attempts to extend the file, the DV advises the CAV as to the new length of the file and, in return, the CAV grants a new ticket book reflective of that new file length. In particular, the CAV invalidates all outstanding ticket books to all DVs, updates the length of the file and returns the new ticket book to the advising DV. Since no other valid ticket books are in use, each DV must poll the CAV for an updated ticket book before servicing a new I/O operation. The new ticket book 1900 has tickets 1920 reflective of higher mtimes, and includes the correct new file length 1925.
For example, assume DV1 services stripes 0 to 2 megabytes of a file, DV2 services stripes 2 to 4 megabytes of the file and the file is currently exactly 2 megabytes in length. While a client issues read operation requests to DV1 for the file, another client decides to extend the length of the file, so it issues a write operation request to DV2. However, DV1 is not notified about that write operation and, as such, its ticket book is not modified or discarded. Therefore, DV1 continues to service read operations to the file as if the length of the file has not increased. The third rule states that if a client tries to span EOF as perceived by a DV (e.g., DV1), then the DV obtains a new ticket book from the CAV. Notably, the new ticket book includes the current length of the file.
This situation is also relevant for write operations because the CAV 1810 is the authority for the file's length. When a DV needs to extend a file via a write operation, it sends a message to the CAV requesting a change of length for the file. The CAV then returns a new ticket book with a new file length attribute that is sufficient to accommodate the write operation. In sum, if any operation request directed to a DV 1815 exceeds the EOF associated with its current ticket book, the DV does not reject that request but rather contacts the CAV for a new ticket book. This third rule provides a way for the DV to obtain a new ticket book from the CAV.
If the new ticket book is returned and the file length has not changed, then the DV proceeds through normal semantics. If a read operation is performed that attempts to span EOF, then the DV returns the number of bytes it was able to read and fills the rest of the request (buffer) with zeros. If a write operation is performed that attempts to span EOF, the DV advises the CAV as to the new length of the file. The CAV makes the necessary change to the file length and returns a new ticket book reflecting the change in file length up to the end of the write operation.
The fourth rule relates to strong serialization semantics. Specifically, the contents of a ticket book can be trusted (used) if (a) weak serialization semantics are used, in which case this rule does not apply, or (b) strong serialization semantics are used and at least one of two tests is satisfied. The first test is whether the DV has a kinetic token. If so, then the DV can use the current ticket book. If not, then the DV must consider the time on the local machine (e.g., the filer) when the ticket book was requested and the time that the incoming file operation request arrived. If the ticket book was requested after that file operation arrived, then the DV can use the ticket book; otherwise the DV cannot use the ticket book (i.e., the ticket book must be discarded and new one obtained from the CAV).
For example, the DV may have started servicing Read A but encountered a portion of the file for an indirect block it does not yet have and, accordingly, Read A is stalled. Meanwhile DV services Read B. If all three Reads A, B, and C are outstanding at the same time from the client's point of view, there is no guarantee as to the order in which the storage system (i.e., D-blade 350) will service the requests. Note that this also applies to outstanding write operations. The only requirement is that once it chooses an order in which to service the (read and/or write) operations, the DV must return mtimes that are consistent with that order. Since all three operation requests were “in flight” simultaneously, the client provides no guarantees for which one is actually going to reach the D-blade first, much less which one gets serviced first. Client protocols that can handle weak serialization semantics are not concerned about the ordering of the operations in real time versus the ticket book mtimes. As noted, the only requirement is that the order in which the operations are serviced (e.g., Reads B, A, C) is consistent with the mtimes that are returned.
In the case of strong serialization semantics graph 2120, the DV receives a first I/O operation (Read A) and, in response, sends a message to the CAV requesting a new ticket book (TB 1). Meanwhile, two more operations (Reads B and C) are received by the DV. When the new ticket book is returned by the CAV, the DV determines that the ticket book is acceptable for servicing Read A, but cannot be used to service Reads B and C because those latter read operations (B and C) arrived at the D-blade after the DV had requested the new ticket book. Accordingly, the DV sends another message to the CAV requesting another new ticket book (TB 2). If Reads A, B and C had arrived before the new ticket book was requested, then the DV would be able to service all three operations with the ticket book TB 1. The last rule states that the DV can only trust (i.e., use) the ticket book for file operations that arrived before the DV had requested the ticket book. It is preferable to avoid strong serialization semantics because of the latencies involved; accordingly, the invention provides a further optimization, called kinetic tokens.
According to yet another aspect of the invention, a kinetic token 1930 is provided that represents an optimization that enables the storage system to defeat caching behavior at the client and improve the performance of file operations. A kinetic token is a guarantee that every time a client requests the current mtime on the file, the client will receive a higher value than has ever been previously reported. Issuance of a kinetic token by the CAV effectively disables client-side caching because every time the client requests the current mtime, it will receive a higher mtime value. As soon as that guarantee is provided, then the ordering problem for strong serialization semantics is eliminated because the clients no longer caches any data (and, as such, it does not matter what mtime value the storage system returns). The kinetic token provides a guarantee that client-side caching is disabled as clients will always see higher mtimes returned. If the DV receives a request that requires strong semantics, the DV utilizes the technique described above and obtains a new ticket book before servicing the request.
The kinetic token is illustratively granted in recognition that there is a relatively common file access pattern. For example, if many write operations are received by a DV 1815 to a particular file, then the mtimes returned to the clients issuing those operations are constantly being updated (incremented). Recall that every write operation results in updating/incrementing of the mtime on the file. In the case of client-side caching, this pattern of activity forces the client to discard the contents of its cache. The CAV may decide to issue kinetic tokens with the ticket book for heavy write access to the file where the mtimes are changing frequently. That is, frequent activity on the file denotes that the CAV is servicing many requests for ticket books 1900 for the same file, often to one or more DVs. If the CAV heuristically determines that many ticket book requests are received for a particular file, it can choose to issue kinetic tokens for that file. Illustratively, the kinetic token 1930 is implemented as a Boolean flag within the ticket book 1900 that, when asserted, specifies existence of the token.
At some point, the ticket book 1900 with kinetic token 1930 expires or the CAV 1810 may revoke the kinetic token. In order for a DV 1815 to manipulate meta-data on the file, the CAV instructs all DVs 1815 that currently have kinetic tokens 1930 to relinquish those tokens. Illustratively, the CAV broadcasts to the DVs a message instructing them to relinquish their kinetic tokens. Once all the DVs have acknowledged that they have given up their tokens, the CAV may halt automatic incrementing of mtimes on query.
The foregoing description has been directed to particular embodiments of this invention. 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. Specifically, it should be noted that the principles of the present invention may be implemented in non-distributed file systems. Furthermore, while this description has been written in terms of N and D-blades, the teachings of the present invention are equally suitable to systems where the functionality of the N and D-blades are implemented in a single system. Alternately, the functions of the N and D-blades may be distributed among any number of separate systems, wherein each system performs one or more of the functions. Additionally, the procedures, processes and/or modules described herein may be implemented in hardware, software, embodied as a computer-readable medium having program instructions, firmware, or a combination thereof. Also the data structures described herein may include additional fields for storing additional information. 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 invention.
This application is a Continuation-in-Part application of U.S. patent application Ser. No. 10/727,169, entitled METHOD AND APPARATUS FOR DATA STORAGE USING STRIPING, by Michael L. Kazar, et al, filed on Dec. 2, 2003, the contents of which are hereby incorporated by reference.
Number | Name | Date | Kind |
---|---|---|---|
6658417 | Stakutis et al. | Dec 2003 | B1 |
20040139167 | Edsall et al. | Jul 2004 | A1 |
Number | Date | Country |
---|---|---|
10003440 | Jun 1998 | JP |
WO 0007101 | Feb 2000 | WO |
Number | Date | Country | |
---|---|---|---|
Parent | 10727169 | Dec 2003 | US |
Child | 11261007 | US |