Methods and systems for implementing shared disk array management functions

Abstract

Multiple Array Management Functions (AMFs) are connected to multiple redundancy groups over a storage area network (SAN), such as a fiber-channel based SAN. The multiple AMFs share management responsibility of the redundancy groups, each of which typically includes multiple resources spread over multiple disks. The AMFs provide concurrent access to the redundancy groups for associated host systems. When a host requests an AMF to perform an operation on a resource, the AMF synchronizes with the other AMFs sharing control of the redundancy group that includes the resource to be operated on, so as to obtain access to the resource. While performing the operation, the AMF send replication data and state information associated with the resource such that if the AMF fails, any of the other AMFs are able to complete the operation and maintain data reliability and coherency.

Description

BACKGROUND OF THE INVENTION

[0002] The present invention relates in general to systems and methods for eliminating bottlenecks in data storage networks, and in direct server attached storage, and more specifically to systems and methods for implementing dynamically shared redundancy group management between multiple disk array management functions.

[0003] The need for faster communication among computers and data storage systems requires ever faster and more efficient storage networks. In recent years, implementation of clustering techniques and storage area networks (SANs) have greatly improved storage network performance. In a typical storage network, for example, N servers are clustered together for a proportional performance gain, and a SAN (e.g., a Fiber Channel based SAN) is added between the servers and various RAID (“Redundant Array of Inexpensive Disks”) storage systems/arrays. The SAN allows any server to access any storage element. However, in the typical storage network, each RAID system has an associated RAID controller that must be accessed in order to access data stored on that particular RAID system. This can lead to bottlenecks in system performance as the storage managed by a particular RAID controller can only be accessed through that RAID controller. Furthermore, if a controller fails, information maintained in the RAID system managed by the failed controller becomes inaccessible.

[0004] One solution for providing fault tolerance is to include a redundant controller in a master/slave arrangement. The master controller has primary control, and only when the master fails does the slave controller take over. This solution is very inefficient, however, as the slave controller is not used until a failure in the master has occurred. Another solution is to use the master/slave controller architecture, but to split the storage array into two redundancy groups, each of which is controlled by one and only one of the two controllers (each controller is a “master” vis-á-vis the redundancy group it controls). In this manner, both controllers are operational at the same time, thereby improving the efficiency of the system. In the event one controller fails, the other controller assumes control of the failed controller's redundancy group. This solution also prevents “collisions”, which occur, for example, when more than one controller tries to write data to a redundancy group. However, this solution also has some performance drawbacks. For example, the performance in such a master/slave architecture is bound by the speed of the master controller such that performance is not scalable.

[0005] Thus, it is desirable to provide techniques for implementing a peer-to-peer controller architecture solution where system performance is not bound by the speed of a given controller. Further, such a system should provide suitable fault tolerance and performance scalability.

SUMMARY OF THE INVENTION

[0006] The present invention provides such a peer-to-peer controller architecture solution for data storage management. The systems and methods of the present invention implement a novel type of RAID Array Management Function that is useful for building highly scalable disk arrays. In particular, the systems and methods of the present invention provide for sharing redundancy group management between multiple (two or more) Array Management Functions.

[0007] According to the invention, multiple Array Management Functions (AMFs) are connected to multiple redundancy groups over an interconnect medium. In one embodiment, for example, the Array Management Functions are connected to the redundancy groups over any storage area network (SAN), such as a fiber-channel based SAN. The multiple AMFs share management responsibility of the redundancy groups, each of which typically includes multiple resources spread over multiple disks. The AMFs provide concurrent access to the redundancy groups for associated host systems. When a host requests an AMF to perform an operation on a resource, the AMF synchronizes with the other AMFs sharing control of the redundancy group that includes the resource to be operated on, so as to obtain a lock on the resource. While performing the operation, the AMF send replication data and state information associated with the resource such that if the AMF fails, any of the other AMFs are able to complete the operation and maintain data reliability and coherency.

[0008] As used herein, the terms “Array Management Function,” “Redundancy Group,” and “Redundancy Group Management” are defined as set forth in The RAID Advisory Board's (RAB) Handbook on System Storage Technology, 6th edition, the contents of which are herein incorporated by reference for all purposes.

[0009] “Array Management Function” (AMF) generally refers to the body that provides common control and management for one or more disk or tape arrays. An AMF presents the arrays of tapes or disks it controls to the operating environment as one or more virtual disks or tapes. An AMF typically executes in a disk controller, an intelligent host bus adapter or in a host computer. When it executes in a disk controller, an AMF is often referred to as firmware. One or more AMFs can execute in each controller, adapter or host as desired for the particular application.

[0010] “Redundancy Group” generally refers to a collection of p_extents organized by an AMF for the purpose of providing data protection. With one redundancy group, a single type of data protection is used. All the user data storage capacity in a redundancy group is protected by check data stored within the group, and no user data capacity external to a redundancy group is protected by check data within it. Redundancy groups typically include logical entities composed of many resources such as stripes, data blocks, cached data, map tables, configuration tables, state tables, etc.

[0011] “Redundancy Group Management” generally refers to the responsibilities, processes and actions of an AMF associated with a given redundancy group.

[0012] “Coherency” generally refers to the process of maintaining an updated copy of the data on a storage subsystem or in cache on an AMF in an environment where multiple copies of the data exist. According to the invention, updates of the check data within a redundancy group are dynamically coordinated and synchronized between the various AMFs sharing the redundancy group. Such updating is facilitated using coherency and locking/unlocking techniques. Coherency and locking are typically performed as a function of a block, a group of blocks, a stripe or a group of stripes. Locking is performed dynamically using any of a variety of well known or proprietary coherency protocols such as MESI. Additionally, the coherency between the caches associated with a redundancy group and the data contained within the redundancy group is synchronized and maintained.

[0013] According to an aspect of the present invention, a method is provided for use in a data storage network for maintaining data coherency using two or more array management functions (AMFs) that are able to concurrently access a redundancy group, the redundancy group including a plurality of resources. The method typically includes receiving a request from a host, by a first one of the AMFs, to perform a first operation on data stored on a first one of the resources, broadcasting a message from the first AMF to the other AMFs sharing access to the first resource so as to acquire access to the first resource, and performing the first operation on the data by the first AMF.

[0014] According to another aspect of the present invention, a method is provided for use in a data storage network for maintaining data coherency using two or more array management functions (AMFs) that are able to concurrently access a redundancy group, the redundancy group including a plurality of resources. The method typically includes receiving a request from a host, by a first one of the AMFs, to perform a first operation on data stored on a first one of the resources, and determining from the request whether the identified data is shared by one or more of the other AMFs. If it is determined that the identified data is shared by one or more AMFs, the method further typically includes broadcasting a message from the first AMF to the other AMFs sharing access to the first resource so as to acquire access to the first resource, and performing the first operation on the data by the first AMF. If it is determined that the identified data is not shared by one or more AMFs, the method further typically includes performing the first operation on the data by the first AMF.

[0015] According to a further aspect of the present invention, a data storage network is provided that typically includes two or more controllers communicably coupled to a redundancy group, each controller having one or more array management functions (AMFs), wherein two or more of the AMFs are able to concurrently access the redundancy group, the redundancy group including a plurality of resources. Each controller further typically includes means for receiving a request from a host to perform an operation on data stored on a first one of the resources by a first AMF, means for broadcasting a message from the first AMF to other AMFs sharing access to the first resource so as to acquire access to the first resource, and means for performing the first operation on the data by the first AMF.

[0016] According to yet another aspect of the present invention, a data storage network is provided that typically includes two or more controllers communicably coupled to a redundancy group, each controller having one or more array management functions (AMFs), wherein two or more of the AMFs are able to concurrently access the redundancy group, the redundancy group including a plurality of resources. Each controller further typically includes means for receiving a request from a host to perform a first operation on data stored on a first one of the resources by a first AMF, means for determining from the request whether the identified data is shared by one or more of the other AMFs, means for broadcasting a message from the first AMF to the other AMFs sharing access to the first resource so as to acquire access to the first resource if the data is shared, and means for performing the first operation on the data by the first AMF.

[0017] Reference to the remaining portions of the specification, including the drawings and claims, will realize other features and advantages of the present invention. Further features and advantages of the present invention, as well as the structure and operation of various embodiments of the present invention, are described in detail below with respect to the accompanying drawings. In the drawings, like reference numbers indicate identical or functionally similar elements.

BRIEF DESCRIPTION OF THE DRAWINGS

[0018] FIGS. 1 to 7 show exemplary configurations useful for providing data from one or more redundancy groups to one or more host systems using controllers sharing access and control of redundancy groups according to the present invention;

[0019] FIG. 8 shows a multiple controller configuration and the internal configuration of the controllers according to the present invention;

[0020] FIG. 9 shows an operation using a general synchronization sequence according to an embodiment of the present invention;

[0021] FIG. 10 shows an operation using a general replication sequence according to an embodiment of the present invention;

[0022] FIG. 11 a shows the flow for read operations when the redundancy group is in a normal, non-degraded mode, according to an embodiment of the present invention;

[0023] FIG. 11 b shows the flow for read operations when the redundancy group is in a degraded mode, according to an embodiment of the present invention;

[0024] FIG. 12 shows the flow for pipelining the replication of incoming host data according to an embodiment of the present invention;

[0025] FIG. 13 a shows the flow for a write operation when the redundancy group is in a normal, non-degraded mode according to an embodiment of the present invention;

[0026] FIG. 13 b shows the flow for a recovery process when the AMF updating the stripe as shown in FIG. 13a fails before completing the update according to an embodiment of the present invention;

[0027] FIG. 14 a shows the flow for a write operation when the redundancy group is in a degraded (with a failed drive) mode, according to an embodiment of the present invention;

[0028] FIG. 14 b shows the flow for a recovery process when the AMF updating the stripe as shown in FIG. 14a fails before completing the update according to an embodiment of the present invention;

[0029] FIG. 15 shows the flow for a background reconstruction process according to an embodiment of the present invention;

[0030] FIG. 16 shows the general sequence flow for a background expansion process according to an embodiment of the present invention;

[0031] FIGS. 17 a and 17b illustrate AMF communication without, and with, the message gathering techniques of the present invention, respectively;

[0032] FIGS. 18 a illustrates a basic arbitration process where an AMF requests a lock for a particular resource according to the present invention;

[0033] FIG. 18 b illustrates the general process flow of the generalized arbitration process according to the present invention;

[0034] FIG. 19 illustrates a simplified arbitration process between two AMFs in a cluster configuration for a single resource;

[0035] FIG. 20 illustrates exemplary resource arbitration sequences for a cluster including four AMFs according to the present invention;

[0036] FIG. 21 illustrates the general information flow for a typical write operation according to the present invention;

[0037] FIG. 22 illustrates the general information flow for an optimized write operation according to the present invention;

[0038] FIG. 23 illustrates the general information flow for a typical read operation according to the present invention;

[0039] FIG. 24 illustrates a simplified data sharing and coherency process for a typical write operation to be performed between two AMFs in accordance with the present invention;

[0040] FIG. 25 illustrates a simplified data sharing and coherency process for a typical read operation to be performed on a resource in accordance with the present invention wherein the requested data is found in cache; and

[0041] FIG. 26 illustrates a simplified data sharing and coherency process for a typical read operation to be performed on a resource in accordance with the present invention wherein the requested data is found on disk.

DETAILED DESCRIPTION

[0042] The present invention provides for shared redundancy group management (SRGM) between multiple AMFs so that multiple AMFs can simultaneously access the same redundancy group. According to the present invention, distributed synchronization and replication techniques are used to coordinate the activities of all AMFs sharing a redundancy group and to maintain data reliability. Access to any redundancy group can be gained through any controller that includes an AMF that is sharing control of that redundancy group. The AMFs sharing a resource group are therefore peers. Additionally, if a redundancy group is configured for shared access and a controller fails, access to data through the failed controller is blocked, but the data on the redundancy group is still intact, protected from disk failures, and accessible from any other controller that includes an AMF which is sharing that redundancy group. Within a given controller, multiple AMFs may be present, in which case redundancy groups are preferably shared on a group by group basis. That is, some redundancy groups may be shared by a first group of AMFs, other redundancy groups may be shared by a second group of AMFs, and still other redundancy groups may not be shared at all. In preferred aspects, within a given controller there is a one-to-one association between an AMF and a redundancy group; that is an AMF manages only one redundancy group at a time. Alternately, an AMF manages multiple redundancy groups at the same time.

[0043] An AMF typically executes in a PCI bus, host bus adapter, a SAN based switch, or a carrier class network edge switch. Also, it will be realized that the AMF may be located in a number of different networks including the Wide Area Network (WAN), Metropolitan Area Network (MAN), Local Area Network (LAN), Broadcast Area Network (BAN), and Multimedia Area Network (M2AN). Also, it should be appreciated that any of a number of different configurations and combinations of AMFs in the different networks as discussed above may be implemented in accordance with the present invention.

[0044] FIG. 1 shows a basic network configuration according to the present invention. As shown, a plurality of network clients 101 to 10N are communicably coupled with a plurality of servers 201 to 20N, each of which includes a controller 30. (In general, “N” is used herein to indicate an indefinite plurality, so that the number “N” when referred to one component does not necessarily equal the number “N” of a different component. For example, the number of clients 10 does not, but may, equal the number of servers 20 in FIG. 1.) Each network client 10 is coupled to one or more of servers 20 over any of a number of connection schemes as required for the specific application and geographical location relative to servers 20, including, for example, an internet connection, any local area network (LAN) type connection, any wide area network (WAN) type connection, any proprietary network connection, etc. Each controller 30 includes one or more AMFs, and is communicably coupled with the multiple arrays 40 of disk drives 45 over an interconnect medium, such as a storage area network (SAN) 50. Preferably, SAN 50 is a fiber-channel based SAN. However, any SAN type, such as a SCSI-based SAN, or any direct server interconnect such as a direct SCSI or FC connection may be used without departing from the spirit of the invention. Because each controller 30 has direct access to each array 40 over SAN 50, redundancy group management can be shared by all of controllers 30.

[0045] A fiber-channel based SAN is preferred because the fiber-channel standard is an open standard that supports several network topologies including point-to-point, switched fabric, arbitrated loop, and any combination of these topologies. Further, a number of protocols are supported by the fiber-channel standard, including SCSI, Asynchronous Transfer Mode (ATM), Transmission Control Protocol/Internet Protocol (TCP/IP), High Performance Parallel Interface (HiPPI), Intelligent Peripheral Interface (IPI), and others. Moreover, Fiber-channel provides for data transfer speeds of up to 100 MBps (200 MBps duplex) at distances of up to 30 meters over copper cabling and up to 10 kilometers over fiber-optic cabling.

[0046] FIG. 2 shows an example of multiple hosts, each with a controller configured in a switch-based fiber-channel SAN according to the present invention. Each controller 30 is coupled to switches 55 in the SAN through two fiber-channel ports as shown. Thus, each controller 30 is in communication with all other controllers 30 and with disk array 40. Each controller 30 communicates with its host system over a PCI bus 35. Switches 55 are coupled to disk array 40 using the loop topology as shown. However, it will be apparent to one of skill in the art that many loops can be supported through any of a number of switching topologies. In general, the more loops, the greater the data transfer rates that can be supported. The system redundancy as shown in FIG. 2 is N-1 , meaning that given N controllers (301 to 30N) up to N-1 controllers can fail and no data on disk array 40 will be lost when all controllers 30 are configured to share management of array 40. A controller failure for a specific host causes a loss of data availability for the specific host, but not for the entire system. Controller environmental faults, such as power supply failures are protected against in this system configuration because the data from one host system is synchronized to data on the other host systems according to the present invention as will be described in more detail below.

[0047] There is a recovery period associated with a controller failure. This is the time it takes for the surviving controllers to make sure that all critical data is again replicated within the cluster. If another controller fails during this recovery period, then data may be lost.

[0048] FIG. 3 shows an example of multiple controllers and a single host configured in a switch-based fiber-channel SAN according to the present invention. Each controller 30 is coupled to the switches 55 in the SAN through two fiber-channel ports as shown, however, from 1 to N ports may be used as desired for the particular application. Thus, each controller 30 is in communication with all other controllers 30 and with disk array 40 over the fiber-channel SAN. Further, each controller 30 communicates with the host system over one or more PCI buses 35. The controllers 30 are also able to communicate with each over the PCI buses 35. Switches 55 are coupled to disk array 40 using the loop topology as shown. However, it will be apparent to one of skill in the art that many loops can be supported through any of a number of switching topologies. In general, the more loops, the greater the data transfer rates that can be supported. In this configuration, up to N-1 controllers can fail before loss of data availability to the host system. External controller 30N+1 is provided to protect against a failure of the host system. If the host system fails, no data will be lost on array 40 when controller 30N+1 is configured to share management of array 40 with the controllers 301 to 30N of the host system according to the present invention.

[0049] FIG. 4 shows an example of multiple hosts each with multiple controllers configured in a switch-based fiber-channel SAN according to the present invention. Each controller 30 is coupled to the switches 55 in the SAN through two fiber-channel ports as shown, however, from 1 to N ports may be used as desired for the particular application. Thus, each controller 30 is in communication with all other controllers 30 and with disk array 40 over the fiber-channel SAN. Further, each controller 30 communicates with its host system over one or more PCI buses 35. The controllers 30 are also able to communicate with each over the PCI buses 35. Switches 55 are coupled to disk array 40 using the loop topology as shown. However, it will be apparent to one of skill in the art that many loops can be supported through any of a number of switching topologies. In general, the more loops, the greater the data transfer rates that can be supported. In this configuration, redundancy and synchronization exist between two or more controllers 30 within each host system. Where each host system includes N controllers 30, up to N-1 controllers can fail before loss of data availability to the host system. If a host system fails, no data will be lost on array 40 when controllers 30 on other hosts are configured to share management of array 40 with the controllers 30 of the failed host system according to the present invention.

[0050] FIG. 5 shows an example of multiple hosts each with multiple controllers configured in a fiber-channel arbitrated loop (FC-AL) SAN according to the present invention. Each controller 30 is coupled to the loop through the two fiber-channel ports as shown. Thus, each controller 30 is in communication with all other controllers 30 and with disk array 40 over the FC-AL. Further, each controller 30 communicates with its host system over one or more PCI buses 35. In this configuration, redundancy and synchronization exist between two or more controllers within each host system. Where each host system includes N controllers 30, up to N-1 controllers can fail before loss of data availability to the host system. Further, if a host system fails, no data will be lost on array 40 when controllers 30 on other hosts are configured to share management of array 40 with the controllers 30 of the failed host system according to the present invention.

[0051] FIG. 6 shows two independent redundancy groups managed by two independent controllers according to the present invention. Redundancy group A is managed by controller 30A of host system A, and redundancy group B is managed by controller 30B of host system B. Also shown are external host system C and external host system D. According to one embodiment, the FC ports of controllers 30A and 30B function as both device and host channels. This allows each controller 30A or 30B to respond to service requests from its associated PCI bus 35 or from an external FC host such as external host system C, external host system D or another controller such as controller 30B or 30A, respectively. In this manner, redundancy group A is made accessible to host system B and redundancy group B is made accessible to host system A. From the perspective of controller 30A, for example, a request received from host system B to read or write data is treated as if it were received over associated PCI bus 35. Similarly, external hosts systems C and D are able to access data on redundancy groups A and B by issuing read or write commands to the appropriate controller 30 over the fiber-channel SAN. Any number of host systems can be interconnected in this manner. Further, although only a single controller configuration is shown, other configurations can be used, such as the configuration as illustrated in FIG. 4. The use of switches helps isolate the disk arrays for performance scalability.

[0052] FIG. 7 shows an example of an external RAID system including multiple controllers in passive PCI Backplane(s) configured in a switch-based fiber-channel SAN according to the present invention. The controllers 301 to 30N are installed into one or more passive PCI backplanes, and are configured to accept host commands from the FC ports and/or the associated PCI buses. Thus, external servers are able to access data on the various redundancy groups controlled by the controllers 301 to 30N by issuing read or write requests to the appropriate controller 30. Although only one controller 30 is required, performance and redundancy scales as more controllers are added. A FC-AL similar to the configuration shown in FIG. 5 can alternatively be used if the use of switches 55 is undesired or impractical.

[0053] FIG. 8 shows a multiple controller configuration and the internal configuration of controllers 30 according to the present invention. One or more of the controllers 301 to 30N shown in FIG. 8 may be located in a separate host or on passive PCI backplanes. For example, each controller 30 may be located in a separate host system, or each of multiple host systems may include one or more of the controllers 30. PCI host connection 60 provides a connection path for receiving and processing commands from host systems and for providing inter-controller link (ICL) services with other controllers. Fiber-channel (FC) host connection 65 provides a connection means for processing and receiving commands from host systems and for providing ICL services with other controllers. In preferred aspects, each controller includes two physical FC ports (not shown in FIG. 8, but see FIGS. 2 through 7 for examples), both of which are used for disk drive access, receiving and processing host commands and ICL services. It will be apparent to one skilled in the art that each controller can include from 1 to N FC ports as desired for the particular application.

[0054] Each controller 30 includes one or more virtual disk ports 70 each of which provides access to a virtual disk 75. Virtual disks 75 are basically partitions of an array. (A “Redundancy Group” is generally synonymous with “Array”). Each array may be partitioned into as many virtual disks 75 as desired. Each virtual disk is associated and controlled by one or more associated AMFs 80. Many virtual disk ports 70 can exist for the same virtual disk 75, but each must exist on a separate controller. For example, as shown in FIG. 8, virtual disk ports 70YR1 and 70YRN associated with virtual disk YR are provided to hosts on controller 301 and controller 30N, respectively. Each virtual disk port YR provides access to the same virtual disk YR. Virtual disk YR is a partition of array Y, the control and management of which is shared by AMFs 80Y1 and 80YN.

[0055] Virtual disk ports can be added to a controller at any time by creating a virtual disk and assigning an IO port address to the virtual disk port. Generally, a virtual disk must exist before a virtual disk port is created and assigned to it, but the creation of a virtual disk is not coincident with the creation of a virtual disk port. Typically a virtual disk port is created right after the redundancy group is created. Virtual disk ports can then be created at any time thereafter, but the creation of the virtual disk is only done once. Virtual disk ports can also be deleted at any time. All host operations in progress on the port are allowed to complete. While these operations are completing, new host operations are rejected, for example, by returning a not ready_status signal to the host.

[0056] Given that multiple AMFs 80 are present on each controller 30, redundancy groups are shared on a group-by-group basis. As shown in FIG. 8, there are three groups of AMFs, Y, G and T, each sharing an array. Associated AMFs on different controllers synchronize and replicate with each other as will be described in more detail below. For example, as shown in FIG. 8, AMF 80Y1 synchronizes and replicates with AMF 80YN (and with any other AMF associated with array Y, e.g., AMF 80Y2 (not shown)). Similarly, AMF 80G1 synchronizes and replicates with AMF 80GN, and AMF 80T1 synchronizes and replicates with AMF 80TN. Additionally, virtual disk ports on one controller synchronize and replicate with related virtual disk ports on other controllers.

[0057] Distributed Synchronization and Replication

[0058] Synchronization and replication ensures that the operations performed by the different AMFs sharing a redundancy group (array) do not destructively interfere with each other (e.g., “collisions” do not occur). Synchronization requires that any AMF which needs to access a resource associated with a shared redundancy group arbitrate with the other AMFs for access rights (lock) on the resource before using it. Arbitration is accomplished by sending arbitration messages between the AMFs over the PCI and/or FC ICL links.

[0059] FIG. 9 shows a general synchronization sequence for an operation according to an embodiment of the present invention. In step 110, the operation is started. For example, a host may send a request that requires data be written to a particular resource. At step 120, the AMF determines whether it already has a lock on the desired resource. If not, the AMF arbitrates with other AMFs for a lock on the desired resource in step 130. Once a lock has been acquired (or it is determined that the AMF already has the lock), the desired operation is performed on the resource by the AMF in step 140. Once a lock is acquired by an AMF, it is preferably not released until another AMF needs the lock (i.e., another AMF arbitrates for the lock) to help cut shared redundancy group management (SRGM) overhead in many applications. According to one embodiment, a first-come-first-served type arbitration scheme is used, but a priority based, or any other arbitration scheme can be used. As will be described in more detail below with reference to FIGS. 18, 19 and 20, arbitration typically involves making a request to use a resource to a resource controller (typically software, but sometimes hardware based). The resource controller grants access to the resource based on the arbitration algorithm used. If a resource is in use when the request is made, the requester waits until the resource is free and all other requesters ahead of it have been serviced.

[0060] Each AMF is able to execute many types of operations on a redundancy group, including, for example, host reads, host writes, background writes, regeneration's, reconstruction's, online expansion, parity scrubbing, etc. An extended sequence of such operations is a termed a “process”. Examples of processes include reconstructions, online expansion, and parity scrubbing. All AMF operation types require synchronization arbitration in order to proceed. Once an operation on a resource is completed by a given AMF, other AMFs are able to use the resource.

[0061] Synchronization is preferably performed at the operation level as opposed to the process level. That is, for each operation to be executed, the basic synchronization sequence shown in FIG. 9 is performed. For a process wherein some function must be performed on the entire redundancy group (such as a reconstruction), the processing is broken up into a sequence of operations. Where each operation operates on a different resource, arbitration for synchronization of those resources required for one operation is done independently of the resources required for other operations in the process. Using synchronization at the operation level instead of the process level allows AMFs to share resources while a process is in progress. If synchronization were performed at the process level instead of the operation level, some AMFs would have to wait until the entire process is complete before they could use the resources, thereby resulting in host timeouts.

[0062] Replication accommodates AMF failures. Resources and their state information are replicated so that if an AMF fails the resource and its state information is available via another AMF that has a copy of the resource and its state information. During the update (writing) of a certain resource, a copy of the modified resource and/or the resource's operation state is sent to other AMFs sharing the resource. These other AMFs are called replication partners. For example, referring to FIG. 8, AMF 80Y1 and AMF 80YN are replication partners as each share control of Array Y. The replicated information is used by the replication partners to complete the operation in the event that the AMF updating the resource fails during the operation.

[0063] FIG. 10 shows a general replication sequence for an operation according to an embodiment of the present invention. The start of the sequence is the basic synchronization sequence as shown in FIG. 9. In step 210, the operation is started. For example, a host may send a request that requires writing data to a particular resource. In step 220, the AMF determines whether it already has a lock on the desired resource. If not, the AMF arbitrates with other AMFs for a lock on the desired resource in step 320. Once a lock has been acquired the operation can be performed. As shown, the operation performed by the AMF is broken into a number, i, of steps. In step 2401, the replication data and state information associated with the resource and the first operation step is sent to each replication partner. In step 2501, the first step of the operation is performed. As shown in FIG. 10, subsequent operation steps 2502 to 250i are performed in sequence, as are the replication steps 2402 to 240i. In each of replication steps 2401 to 240i, the replication information is sent to the replication partners associated with the given resource. In general, N-1 concurrent AMF failures are accommodated if N copies of a resource and its state information exist within the AMF cluster (i.e., those AMFs sharing the resource), where N is defined as the replication dimension. Thus, replication information is sent to the N-1 replication partners associated with the given resource.

[0064] Replication can be targeted to specific replication groups or specific AMFs. In one embodiment, N-way replication is performed without defined replication groups. In this embodiment, replication takes place with any N-1 AMFs in the cluster that are sharing the resource being replicated. Alternately, replication is performed with N-1 other replication groups. A replication group is a group of AMFs that replicate critical data to AMFs in replication groups other than their own. An example of this is a set of controllers, each controller including one or more AMFs, at one physical site and another set of controllers at another physical site. Another example is a set of controllers inside a host system and another set external to a host. Using replication groups helps ensure that if one group of controllers all fail, the other group(s) have the information necessary to maintain data reliability.

[0065] The replication can be targeted to specific replication groups or specific AMFs. A given replication group preferably replicates with any AMF outside of the replicating AMFs replication group. However, the set of replication groups to which a given replication group replicates may be specified by an operator. Also, synchronous replication is the preferred replication mode. In the synchronous replication mode, completion status information for an operation is returned to a host after all replication targets have received the replication data. Alternate replication modes include asynchronous replication and pseudo-synchronous replication. In the asynchronous replication mode, completion status information for an operation is returned to a host before replication data has been transmitted. In the pseudo-synchronous replication mode, completion status information for an operation is returned to a host after the replication data has been transmitted, but before all replication targets have acknowledged receiving the data.

[0066] According to the present invention, multiple AMFs are able to read a resource concurrently. That is, multiple read locks can be outstanding at any time. However, only one AMF at a time writing to the resource is accommodated. Allowing multiple AMFs to read a resource at the same time significantly improves read performance. If an AMF is writing to the resource, no other AMF is permitted to read the resource. The synchronization protocol allows and enforces this aspect.

[0067] FIG. 11 a shows the general sequence flow for a read operation when the redundancy group (RG) is in a normal, non-degraded mode according to an embodiment of the present invention. “Non-degraded” generally refers to the case where all drives in the redundancy group are operational, whereas “degraded” generally refers to the case where one or more drives in the redundancy group have failed. In step 310, the read operation is started. For example, the AMF receives a request from a host to read a particular resource. In step 320, a lock on the particular resource is required. This is basically the same as steps 120 and 130 of FIG. 9. In preferred aspects, multiple read locks can be outstanding. This enables multiple AMFs to read a resource concurrently.

[0068] FIG. 11 b shows the general sequence flow for a read operation when the redundancy group (RG) is in a degraded mode according to an embodiment of the present invention. In step 310, the read operation is started. For example, the AMF receives a request from a host to read a particular resource. In step 320, a lock on the particular resource is required. Once the lock is acquired, in the degraded read mode, the AMF reads the data and parity from the particular resource at step 340, and regenerates any missing data at step 350. In step 360, the data (regenerated) is transferred to the host that issued the read request.

[0069] FIG. 12 shows the general sequence flow for replicating incoming host data in a pipelined manner for a write operation according to an embodiment of the present invention. Pipelining of the replication data helps to minimize replication latency. In step 410, the operation is started. For example, a host issues a write command to write one or more blocks of data to one or more resources. In step 420, the host command is received from the host and parsed. The host command is processed as a sequence of data block sets. In step 430, the appropriate lock for the starting set of blocks is acquired. In step 440, the starting block set is transferred to the AMF from the host. In step 450, the block set replication is started for the starting set of blocks. The AMF does not wait for the block set replication to complete; the AMF immediately determines whether any more sets of blocks need to be processed at step 460. If so, the AMF immediately starts acquiring the appropriate lock to get the next set of blocks in step 430, and repeats steps 440, 450 and 460 for the next block set. If all block sets have been received and processed, the AMF waits for all replication operations to complete in step 470. When each operation is complete the AMF sends status to the host in step 480.

[0070] FIG. 13 a shows the general sequence flow for a write operation when the redundancy group (RG) is in a normal, non-degraded mode according to an embodiment of the present invention. In step 510, the operation is started. For example, a host issues a write command to write data to a resource. Step 520 is the process of acquiring the synchronization lock for the resource required as shown in FIG. 9. In this example the resource is a stripe write lock, but it may also be any other lock as required by the particular operation. After the stripe write lock is acquired, the AMF reads the old data and parity from the RG's disks in step 530.

[0071] Concurrent with the disk read operation of step 530, the AMF sends a state notification signal to its replication partners for this resource, in step 540. The replication partners include all other AMFs to which this AMF replicates state information and data for the particular resource. The number of replication partners is equal to N-1 where N is the replication dimension. Preferably, the replication dimension N is from 1 to 8, but N may be any number as desired. The state notification signal is a ‘begin update’ type signal, which tells the replication partners that a stripe update has started. The replication partners need to know this information because they will be responsible for cleaning up in the event the writing AMF fails before completing the operation.

[0072] Once the ‘begin update’ notification step 530 and the read of old data and parity step 540 are complete, the AMF writes the new data to the RG member disks in step 550. Concurrent with the new data write step 550 is the generation of the new parity in step 560. Thereafter, in step 570 the AMF writes the new parity to the RG member disks. Once the parity write operation is complete, the AMF sends an ‘end update’ notification to its replication partners in step 580. Upon receiving this notification, the replication partners release their state data associated with the stripe update.

[0073] FIG. 13 b shows the general sequence flow for a recovery process when the AMF updating the stripe as shown in FIG. 13a fails before completing the update according to an embodiment of the present invention. When the AMF updating the stripe fails before completing the update, the replication partners assume the responsibility of recovering from the failed update operation. In step 610, the recovery operation begins when one or more of the replication partner AMFs either detects a failure or is notified of a failure. The replication partners arbitrate for ownership of the stripe lock in step 620. The AMF that wins the arbitration (the recovery AMF) is responsible for executing recovery of the failed update operation.

[0074] Failure notification typically comes from the inter-controller link (ICL) component. If a controller fails, the AMFs on that controller lose communication with the other AMFs they were sharing the redundancy group with. The ICL periodically sends a ‘ping’ message to all the other AMFs it is sharing the redundancy group with. If any of these AMFs fails to respond to the ping message, then the AMF that sent the ping message assumes the AMF has failed and begins recovery action. Recovery is also triggered if the ICL encounters a transport failure when sending synchronization or replication messages to the destination AMF.

[0075] The recovery process includes 2 basic steps: recalculation of the stripe parity and rewriting the data. In step 630, the recovery AMF reads all the data for the stripe segments affected by the failed update operation. Concurrent with the data read step 630, the recovery AMF assigns one or more new replication partners and sends a ‘begin update’ notification to all of its replication partners in step 640. When the data read operation is complete, the recovery AMF generates new parity in step 650. This new parity calculation does not include the new data. It is simply a regeneration of parity for data on the RG member disks.

[0076] Once the parity generation step 650 and ‘begin update’ notification step 640 are complete, the recovery AMF writes the new parity to RG member disks in step 660. After the parity write step 660 is complete, the recovery AMF sends an ‘end update’ notification to the replication partners in step 670. After some time interval, the cache write back scheduling algorithm causes one of the replication partners to write the new data to the RG member disks in step 680, which is a normal (non-recovery mode) stripe update operation as shown in FIG. 13a. The caching functionality is a part of the AMF.

[0077] FIG. 14 a shows the flow for a write operation when the redundancy group (RG) is in a degraded (with a failed drive) mode, according to an embodiment of the present invention. This sequence is similar to that of the non-degraded case shown in FIG. 13a, with the inclusion of regeneration step 744 and replication step 746 as will be described below.

[0078] In step 710, the operation is started. For example, a host issues a write command to write data to a resource. Step 720 is the process of acquiring the synchronization lock for the resource required as shown in FIG. 9. In this example the resource is a stripe write lock, but it may also be any other lock as required by the particular operation. After the stripe write lock is acquired, the AMF reads the old data and parity from the RG's disks in step 730.

[0079] Concurrent with the disk read operation of step 730, the AMF sends a state notification signal to its replication partners for this resource, in step 740. The replication partners include all other AMFs to which this AMF replicates state information and data for the particular resource. The state notification signal is a ‘begin update’ type signal, which tells the replication partners that a stripe update has started. The replication partners need to know this information because they will be responsible for cleaning up in the event the writing AMF fails before completing the operation. Once the read of old data and parity step 540 is complete, the AMF regenerates the data that was on the failed disk in step 744. In step 746, the old data, including regenerated data, is replicated to the replication partners. Replication of this data to the replication partners is necessary for recovery in the event the updating AMF fails before completing the operation.

[0080] Once replication step 746 is complete, the new data is written to the RG member disks in step 750. Concurrent with the new data write step 750 is the generation of the new parity in step 760. Thereafter, in step 770 the AMF writes the new parity to the RG member disks. Once the parity write operation is complete, the AMF sends an ‘end update’ notification to its replication partners in step 780. Upon receiving this notification, the replication partners release their state data associated with the stripe update.

[0081] FIG. 14 b shows the general sequence flow for a recovery process when the AMF updating the stripe as shown in FIG. 14a fails before completing the update according to an embodiment of the present invention. This scenario differs from the non-degraded recovery case shown in FIG. 13b in that the recovery AMF uses the replicated old data to reconstruct the RG disk segments that were affected by the updating AMF's failure. When the AMF updating the stripe fails before completing the update, the replication partners assume the responsibility of recovering from the failed update operation. In step 810, the recovery operation begins when one or more of the replication partner AMFs either detects a failure or is notified of a failure, for example by a host. The replication partners arbitrate for ownership of the stripe lock in step 820. The AMF that wins the arbitration (the recovery AMF) is responsible for executing recovery of the failed update operation.

[0082] In step 830, new parity is generated from the old data supplied by replication step 746 of FIG. 14a. Concurrent with step 830, the recovery AMF assigns one or more new replication partners and sends a ‘begin update’ notification to all of its replication partners in step 840. In step 850, the old data is written to the disks. In step 860, the replication partners are informed that the old data has been written back to the disks. The replication partners can now discard their copies of the old data. After this event, the recovery sequence is the same as for the non-degraded recovery sequence. Specifically, the new parity to written to the RG member disks in step 870. After the parity write step 870 is complete, the recovery AMF sends an ‘end update’ notification to the replication partners in step 880. After some time interval, the cache write back scheduling algorithm causes one of the replication partners to write the new data to the RG member disks in step 890, which is a normal (non-recovery mode) stripe update operation as shown in FIG. 13a.

[0083] Distributed Reconstruction

[0084] FIG. 15 shows the general sequence flow for a background reconstruction process, according to an embodiment of the present invention. Each operation is started in step 910, and the appropriate lock is acquired in step 920, in this case a stripe lock. In step 930, the AMF reads the data and parity for the stripe. In step 940, the AMF regenerates missing data, and in step 950 writes the data to the replacement disk. After writing the data, the AMF updates the map table to reflect the fact that blocks that originally mapped to the failed disk now map to the replacement disk in step 960. The map table maps host logical blocks to RG member disks and blocks on the disks. In step 970, it is determined whether there are any more stripes to reconstruct. This sequence repeats as long as there are more stripes that need to be reconstructed onto the replacement disk.

[0085] One advantage of the shared redundancy group management (SRGM) techniques of the present invention is that SRGM allows for distributing reconstruction processing to all AMFs sharing a RG. This results in faster reconstruction times and a general reduction in the increased response times normally encountered by a host during reconstruction. Distributed reconstruction is accomplished by having a subset (or all) of the AMFs sharing a RG coordinate for which stripes each will be responsible for reconstructing. If any of these AMFs fail or shutdown during reconstruction, the remaining AMFs re-arbitrate reconstruction responsibilities. For example, suppose there are N AMFs sharing a redundancy group that needs reconstruction. These AMFs talk to each other (by sending messages) and determine which ones are to participate in the reconstruction, e.g., a subset of N, denote by M. These M AMFs determine reconstruction responsibilities by determining which AMFs will reconstruct which stripe(s). This can be determined by any algorithm. In one embodiment of the invention, for example, each AMFi reconstructs stripe(s) Mod(S/i), where S is the stripe number.

[0086] Distributed RG Expansion

[0087] Just as reconstruction is preferably distributed among a subset of the AMFs sharing a RG, so is RG expansion. RG expansion is the addition of drive members to an existing RG. A unique advantage of SRGM is that it allows expansion processing to be distributed to all AMFs sharing a RG. This results in faster expansion times and a reduction in the increased response times normally encountered by a host during expansion.

[0088] Distributed expansion is accomplished by having a subset (or all) of the AMFs sharing a RG arbitrate for which stripes they will be responsible for expanding. If any of these AMFs fail or shutdown during expansion then the remaining AMFs re-arbitrate expansion responsibilities. For example, suppose there are N AMFs sharing a redundancy group that needs expansion. These AMFs talk to each other (by sending messages) and determine which ones are to participate in the expansion, e.g., a subset of N, denoted by M. These M AMFs determine expansion responsibilities by determining which AMFs will expand which stripe(s). This can be determined by any algorithm. In one embodiment of the invention, for example, each AMFi expands stripe(s) Mod(S/i), where S is the stripe number.

[0089] FIG. 16 shows the general sequence flow for a background expansion process according to an embodiment of the present invention. The process is started in step 1010, and the appropriate stripe lock is acquired in step 1020. The expansion case is different from the preceding examples in that multiple locks must be acquired. An expansion operation will involve 2 or more stripes. One stripe is the expansion stripe whose stripe width will be increased from W to W+N. The other stripes involved are stripes containing host data that will be migrated from those stripes to the expanded stripe.

[0090] In step 1030, the data on the stripe to be expanded is read. In step 1040, the data is replicated so that if the operation fails before completion the replication partners will be able to clean up after the failure and continue the expansion process. Concurrent with steps 1030 and 1040, the source data stripe containing data to be migrated to the expanded stripe is read in step 1045. After steps 1040 and 1045 are complete, the AMF notifies its replication partners that it is beginning the expansion stripe update in step 1050. Concurrently, in step 1055, the AMF generates parity information for the expanded stripe.

[0091] After the notify begin update has completed, the data for the expanded stripe is written to the disks in step 1060. Once the parity generation step 1055 and the notify begin update step 1050 are complete, the parity is written to the disks in step 1070. Once steps 1060 and 1070 are complete, the AMF notifies its replication partners that the update is complete in step 1080. The replication partners then update their map tables to reflect the increased stripe width and migrated host data. They also discard the data replicated in step 1040. The map table maps host logical blocks to RG member disks and blocks on the disks.

[0092] In step 1090, it is determined whether any more stripes are to be expanded by the AMF. If so, the sequence is repeated. This sequence repeats as long as there are more stripes that need to be expanded to utilize the capacity of the new RG member disks. Note that this is process—what makes it a process is the looping that results in step 1090. Steps 1020 through 1090 comprise an operation.

[0093] Message Gathering

[0094] One issue of concern with SRGM is the processing overhead and IO channel loading associated with the synchronization and replication used to support SRGM. To facilitate the synchronization and replication techniques, ICL (Inter-Controller Link) message gathering is preferably used. Message gathering is, generally, an algorithm that combines many small messages destined for a particular cluster node (i.e. a controller, in which may reside many AMFs) into one big message packet and sends it as one message to the particular node. This dramatically reduces processing overhead and IO channel loading, and contrasts with the approach of sending individual messages to a cluster node.

[0095] FIG. 17 a illustrates AMF communication without the message gathering techniques of the present invention. As shown, a collection of AMFs 1100 and an Inter-Controller Link (ICL) entity 1105 compose a SRGM node 1110. A node is typically a hardware entity such as a controller. ICL 1105 is a software entity that is responsible for routing synchronization and replication messages 1120 from one AMF to another. As shown in FIG. 17a, only one of many similar nodes is shown as being connected to the SAN 1130.

[0096] The AMFs 1100 within node 1110 are sending and receiving synchronization and replication messages with other AMFs on other nodes that share the same redundancy group. Each AMF within node 1110 generates independent streams of synchronization and replication messages, all destined for one or more other nodes on SAN 1130. The messages being sent or received by a particular AMF are independent of the messages being sent or received by other AMFs on the same node. As shown in FIG. 17a, three AMFs 1100 are sending a total of nine messages 1140 to AMFs on other nodes. Without message gathering, ICL 1105 has to send nine messages to other nodes. Also, without message gathering, all synchronization and replication messages generated by all AMFs within a SAN node are processed and sent through the SAN individually. Each message takes a fixed amount of processing overhead, regardless of size.

[0097] FIG. 17 b illustrates AMF communication with the message gathering techniques of the present invention. Message gathering is where many smaller messages destined for a particular node are packed together to form one larger message. This larger message can be sent over SAN 1130 as one message and then unpacked on the receiving node back into the individual messages. For example as shown, the nine messages 1120 are destined for three different nodes. In this example, then, if message gathering is used, ICL 1105 only needs to send three messages 1150—one for each node (not counting itself). ICL 1105 takes on the responsibility of packing and unpacking individual AMF messages.

[0098] Arbitration and Distributed Resource Arbitration

[0099] FIG. 18 a illustrates a basic arbitration process where an AMF requests a lock for a particular resource according to the present invention. AMF 1200 and AMF 1210 each request a lock on a particular resource, and the lock requests are queued in an arbitration queue 1205. The arbitration process for an AMF begins when a request is placed in arbitration queue 1205. The requests are processed in some order such that all requests are satisfied in priority order. The request queue priority is established through any well known algorithm (e.g. FIFO, LIFO). Each requesting AMF must wait until its request is processed to obtain the lock. Each AMF obtains a lock on the resource at successful completion of the arbitration process. An AMF fails to lock the resource if arbitration fails.

[0100] FIG. 18 b illustrates the general process flow of the generalized arbitration process according to the present invention. Arbitration involves coordinating the resource lock transfer between two AMFs: the requesting AMF 1225 and the AMF 1230 with the resource lock. AMF 1225 sends a Request Lock message to an arbitrator 1220 (the entity performing the arbitration process), which queues the message until a time defined by its priority algorithm. When appropriate, arbitrator 1220 processes the request by issuing a Release Lock message to AMF 1230 which currently has the resource lock. Once it has reached a point where it can safely discontinue use of the resource, AMF 1230 releases the lock and notifies arbitrator 1220 that the lock is released. Arbitrator 1220 then signals requesting AMF 1225 that it has been granted the resource lock. AMF 1225 continues to hold the lock until arbitrator 1220 calls for it to release the resource.

[0101] Optimizations of the arbitration queue are possible when one or more AMFs request a read lock for a particular resource. In preferred aspects, the arbitration process simultaneously grants read locks in these situations, as long as command ordering is preserved.

[0102] An AMF (or controller) manages the arbitration process for a resource within the redundancy group cluster. This AMF is known as the resource arbitrator. Assignment of the arbitrator for a specific resource can be accomplished using any of multiple methods (e.g. single arbitrator for all resources, load balancing assignment, etc.). The preferred methods for arbitration assignment according to the present invention are based on the number of controllers and the resource range. For cluster configurations with one or two AMFs, the assigned arbitrator is the last AMF with a Write Lock. For cluster configurations with greater than two AMFs, arbitrator assignment is based upon a fixed mapping algorithm based upon the number of AMFs and the resource location (e.g. stripe ID, block ID, etc.).

[0103] FIG. 19 illustrates a simplified arbitration process between two AMFs in a cluster configuration for a single resource. First AMF 1300 (AMF #1) issues a Read Lock request 1320 for a resource to second AMF 1310 (AMF #2), which currently has a Write Lock on the resource. AMF#2 issues a Grant Lock (read) message 1330 to AMF #1 indicating that a resource lock has been granted. AMF #1 now has read access to the resource. The sequence continues when AMF #1 issues a Write Lock request 1340 to AMF #2. AMF #2 responds with a Grant Lock (write) message 1350. Continuing, AMF #1 issues a Read Lock request 1370, and since AMF #1 already has a Write Lock, it handles its own arbitration and demotes the Write Lock to a Read Lock. Next, AMF #2 has no locks on the resource at this time, so it does not need to be notified. AMF #2 issues a Read Lock request 1370 to AMF #1, which responds immediately with a Grant Lock (read) message 1380 since the resource supports multiple read locks. For the two-AMF cluster configuration, arbitration with the other AMF is not required when the AMF with a Write Lock is requesting a Read Lock. In all other cases, the second AMF must arbitrate the lock.

[0104] FIG. 20 illustrates exemplary resource arbitration sequences for a cluster including four AMFs according to the present invention. The preferred arbitrator assignment method for clusters containing three or more AMFs is to select the arbitrator using a fixed mapping algorithm. This has the effect of permanently associating an arbitrator with a single AMF. In the event of AMF resource arbitration failure, the resource arbitrator is reassigned according to the mapping algorithm.

[0105] Referring to the FIG. 20, First AMF 1400 (AMF #1) issues a write lock request 1420 to the resource X arbitrator on third AMF 1410 (AMF #3). The arbitrator on AMF #3 issues a release lock request 1422 to second AMF 1405 (AMF #2) to release its lock on the resource X. While the lock request 1420 for resource X is processing, AMF #1 issues a read lock request 1424 for resource Y. Fourth AMF 1415 (AMF #4) is the assigned arbitrator for resource Y. AMF #4 immediately grants a read lock 1426 since no other AMFs currently have a lock. Next, AMF #4 issues a write lock request 1428 for resource X, which is queued by the arbitrator on AMF #3 since it is currently processing write lock request 1420.

[0106] AMF #2 sends a lock released message 1430 to AMF #3, which sends a grant lock (write) message 1432 to AMF #1. Embedded within grant lock message 1432 is a flag indicating that AMF #1 should release the lock when finished. This optimization eliminates the need for AMF #3 to send a release lock message to AMF #1. When finished with resource X, AMF #1 sends a lock released message 1434 to AMF #3, which sends a grant lock message 1436 to AMF #4 (which is next in the queue for a write lock on resource X).

[0107] The sequence beginning with request lock (read) message 1440 shows a multiple read lock condition. With the reception of the grant lock (read) message 1442, both AMF #2 and AMF #1 have simultaneous read locks on resource Y. The write lock request 1444 sent by AMF #3 causes AMF #4 to issue release lock messages 1446 and 1448 to AMF #2 and AMF #1, respectively. This results in both lock-released message 1450 and lock-released message 1852 being sent to AMF #4. Prior to AMF #4 granting a lock to AMF #3, AMF #1 sends a request read lock message 1454 which is queued by AMF #4. AMF #3 receives the grant write lock message 1456 for resource Y which contains a flag indicating that it should release the lock when complete. AMF #3 issues a lock released message 1458 when done with the resource Y. AMF #4 then issues a grant lock (read) message 1460 notifying AMF #1 that it has obtained a read lock on resource Y.

[0108] One benefit of this method for lock management is that the number of messages sent between the AMFs is only a function of the number of locks outstanding. It is not dependent upon the number of AMFs in the cluster configuration. Extending the first method, for example as illustrated in FIG. 19, for the two-controller cluster configuration to support more than three controllers will cause the number of arbitration messages to grow with the number of AMFs in the cluster configuration. Other methods for arbitration optimization are possible, but these two methods present the preferred embodiment for the current invention.

[0109] Resource Prefetching

[0110] In the preferred embodiment for the current invention, resource arbitration is also optimized through the use of lock prefetch. An AMF can specify additional prefetch resources when arbitrating for a lock. If all or some of the prefetch resources are not locked, the Arbiter will lock them for the AMF as well. Thus, when the AMF requests the lock on these prefetched resources, (at some later time) it can quickly gain the lock (since it already had it).

[0111] Data Sharing and Coherency

[0112] “Sharing Partners” generally refers to a plurality of AMFs that each cache data blocks for the purpose of increased bandwidth. Because there are multiple copies of the data blocks, each copy of the data block must be kept coherent, i.e., each copy contains the most recent value of the data block.

[0113] “Replication Partner” generally refers to a plurality of AMFs that each contain a copy of a cached data block for the purpose of recovery in the event that the cached copy of the data is lost. These multiple copies must be identical so that any one copy of the data may be used for recovery purposes.

[0114] Note that the replication partners may be a different from the sharing partners.

[0115] FIG. 21 illustrates the general information flow for a typical write operation, in accordance with an embodiment of the present invention. In step 2000 the write operation is started. For example, a host issues a write command to write a particular data item to a redundancy group. As mentioned previously, the redundancy group is typically comprised of a plurality of resources such as disks. The AMF receiving the write request proceeds to perform a write invalidate request broadcast to all of the sharing partners in step 2020. The broadcast (step 2020) is used to invalidate and serve as notification to the other N-1 AMFs that a cached copy of the data item (if any) on the AMFs is stale, or out of date as of the subsequent local cache read operation performed on each AMF having a local cached copy of the data item.

[0116] During step 2040, the AMF decides whether to perform a write thru cache operation or whether to perform a write back cache operation. During a typical write-thru cache operation, the data to be written from the host is written directly to disk. During a typical write-back cache operation, the data to be written from the host is not written directly to disk, but rather, the data is cached on the AMF and a write completion acknowledgment is issued to the host before the data is written to disk. The data is subsequently written to disk at a time after the issuance of the write completion acknowledgment. Subsequent to step 2040, the data is written to disk during step 2050. In a write-back operation, the data is replicated to the other N-1 replication partners in the replication group in step 2060 (where N stands for the number of AMFs of the replication group receiving a copy of the data item).

[0117] In step 2070, a completion operation notification is issued to the host indicating that the write operation is completed. The following conditions must be satisfied before the write can be acknowledged back to the host. Firstly, the invalidate broadcast must be completed, which means that all broadcast recipients have acknowledged that they have invalidated their copy of the data (if it existed). Secondly, the data must be committed, which means that it must have been transferred successfully from the host and, in the case where an alternate buffer was used because the data was still being read, the alternate buffer has become the primary data buffer. Thirdly, the data must be deemed to be “safe”, meaning that it has either been written to disk or replicated to the replication group.

[0118] As shown in FIG. 22, an optimization to the general coherency algorithm of FIG. 21 may be achieved by determining whether the data item is shared in step 2010. If, the data is shared, a write invalidate request is broadcast in step 2020 to the other sharing partners sharing access to the redundancy group containing that data item item. If the data item is not shared, the broadcast (step 2020) is unnecessary. Normally, the write invalidate request is issued to all members (sharing partners) who might be sharing the data, which includes any AMFs that have accessed data on a Redundancy Group (or virtual partition). These sharing partners may be different from the replication partners (of step 2060). The Replication Partners are a group of AMFs that receive a copy of the written data item for fault-tolerance reasons such that if any of the N-1 replication partners fail, or go away for whatever reason, a copy of the data item is available to the requesting AMF, which AMF is communicating with the other N-1 replication partners.

[0119] In one embodiment of the invention, the scope of the broadcast (step 2020) is minimized by applying optimizations to the coherency algorithm. For example, intelligent decisions about which AMFs to send the invalidate request may be determined depending upon the type of operation performed. For example, if an AMF performed a write operation on the data then it must have invalidated all other copies of the data on other AMFs. Further, if any other AMF wanted to perform a read operation on the data, it must have communicated with the AMF that wrote the data. Thus, by remembering if the data has been shared since the last write operation, it is possible to eliminate the broadcast of step 2020 by the determination step 2010.

[0120] Remembering if the data has been shared typically requires a mechanism for providing a history of which AMF(s) last updated a particular data item item. Such a mechanism of providing and updating this information is preferably provided in a table or directory-based scheme (e.g., in the cache of one or more of the AMFs). A simple optimization is only to remember that the data was requested by another AMF and not to remember which AMF requested the data. This eliminates the need for a table. If the data has not been shared, no invalidate broadcast is needed. A further optimization is to remember exactly the AMF(s) that requested the data, and send the invalidate message to only those AMFs, since these AMFs are the only ones who have copies of the data. Again, remembering requires a sharing history table.

[0121] As shown in both FIGS. 21 and 22, the transfer of the data item required to be written by the host is transferred to the designated AMF performing the operation. The transfer of the data item may occur concurrent with the invalidate broadcast. In one embodiment, the transfer may occur while the original data item is still being transferred to the host by a previous read operation (a separate buffer is used for the new data). The only precondition before the transfer from the host can begin is that all local writes have completed.

[0122] FIG. 23 illustrates the general information flow for a typical read operation according to the present invention. In step 2100, the read operation is started. For example, a host issues a request to read data from a particular resource. In step 2110, the AMF receiving the read request checks for a local copy of the data in cache. If a local copy of the data is found in cache, the data is provided to the host directly by the AMF receiving the read request during in step 2110. However, if a local copy of the data is not found in the cache (cache miss) of the AMF receiving the read request, then a shared read request is broadcast to all other sharing partners in step 2120 by the AMF. If a local copy of the requested data is found in the cache of one of the other N-1 AMF sharing partners, then in step 2135 a copy of the data is provided to the requesting AMF which caches the copy locally. The requesting AMF provides the data to the host in step 2136 and the operation is completed. However, if no copy of the data is found in the cache of a replication partner, then the data is read directly from disk in step 2130 and provided to the host. In step 2140, a completion operation notification is issued to the host indicating that the read operation is completed.

[0123] FIG. 24 illustrates a simplified timing diagram for the data sharing and coherency process for a typical write operation to be performed between two or more AMFs in accordance with the present invention. The case of data sharing and coherency between four (4) AMFs will be discussed first with additional reference to FIG. 21. The same methodology of data sharing as set forth herein may be applied to N AMFs comprising the sharing partners. First AMF 2200 (AMF#1) issues a write invalidate request 2210 to all of the AMFs in the sharing group (AMF#2-AMF#4). The AMFs that receive the request each invalidate their local copy of the data item specified in the invalidate request. Each AMF in the sharing group issue an invalidate acknowledgment message 2230 based on the invalidate request 2210 issued from AMF #1. The data to be written from the host (e.g., during step 2030 of the write process of FIG. 21) is received from the host by AMF#1. Note that the invalidate operation (different from the invalidate message) and subsequent invalidate acknowledgement is performed on multiple AMFs that are different from the AMF that issued the invalidate message. Thus, their operations can be performed separately and simultaneously (concurrently) to the actions performed on the AMF that issued the request.

[0124] For the purposes of this example, assume that AMF#2 is a replication partner of AMF#1, and that write-back mode is in effect. After receiving the data, AMF#1 sends a replication message 2240 to AMF#2 (the corresponding replication partner). Continuing, AMF #2 issues a replication acknowledgment message 2250 which is received by AMF#1. The replication acknowledgment message 2250 provides state information to AMF #1 and indicates that AMF#2 has received the replication data. The write operation proceeds and the data is written to disk (e.g., in step 2050 of FIG. 21).

[0125] Concurrent with performing the invalidate request 2210, the replication message 2240 is sent to all of the N-1 replication partners 2220 which are sharing the resource with AMF #1. Each of the N-1 replication partners 2220 send a replication acknowledgment request 2260 in response to the replication request 2240. Accordingly, if AMF #1 fails before completing the operation, any of the N-1 AMFs 2220 will be able to complete the request by reconstructing the state information. As mentioned previously, the replication of actual data (for fault-tolerance aspects) is distinguishable from replicating state information as shown by step 2240 in FIG. 24. For example, AMF #1 which is responsible for receiving and performing (or outsourcing in the present example) the request to the other N-1 AMFs may decide to perform a replication operation of the actual data across the other N-1 AMFs. Accordingly, other cached copies of the requested data may be present within the cache of some or all of the other N-1 AMFs such that if any of the AMFs physically goes away, another copy of the data is found (in cache) on another AMF. Of course, it is assumed that a copy of the data is also present on disk, however, going back to disk to retrieve the data is generally not advisable (due to the high latency).

[0126] FIG. 25 illustrates a simplified timing diagram for a data sharing and coherency process for a typical read operation to be performed on a resource for the case when the data is found in cache on an AMF. The case of data sharing and coherency between only four (4) AMFs will be discussed with additional reference to FIG. 23. The same methodology of data sharing as set forth herein may be applied to N AMFs comprising the sharing partners. First, AMF 2300 (AMF#1) issues a shared data broadcast request 2310 (step 2120 in FIG. 23) to all sharing partners (AMF#2-AMF#4).

[0127] Each of the sharing partners (AMF #2 to AMF #4) 2350 issues an individual acknowledgment request 2330 to the requesting AMF. If AMF #2 has the requested data in cache, a positive acknowledgment response 2330 would be issued. If any of the other AMFs 2350 receiving the broadcast has a copy of the data, then a positive acknowledgment 2370 is sent from each of those sharing partners 2350, since accordingly none of the AMFs need to know the state of the other AMFs. In accordance with this principle, AMF #1 may receive multiple positive acknowledgment requests 2370 from sharing partners 2350. In the event that any of the sharing partners 2350 fails to deliver the requested data to AMF #12300, any of the other N-1 sharing partners 2350 will be able to finish the operation of delivering the data to AMF #1.

[0128] In one embodiment of the invention, a copy of the data is maintained on multiple AMFs 2350 such that multiple positive acknowledgments are sent from each of these AMFs 2350 to the AMF 2300 requesting the data. In the event of a failure of any M(≦N) AMFs 2350, a positive acknowledgment is received from N-M AMFs 2350 that have a copy of the data. This provides a contract between the requesting AMF 2300 and the AMF 2350 sharing partners (which may be different than the replication partners) that the relevant coherency information will be maintained. Also, it may be the case that one of the AMFs 2350 that contains the requested data is busy and cannot reply immediately. Having multiple copies of the positive acknowledgment reply allows the requesting AMF 2300 to use the first one it receives and simply ignore the others. Preferably, each of the AMFs 2320 includes a table (directory) listing which AMFs have the requested data in cache memory. As an optimization for the broadcasting methodology, the requesting AMF 2300 may go directly to the AMF(s) which have the requested data by searching the table. Accordingly, this optimization is a hybrid of the directory-less cache coherency protocol, in accordance with the present invention, and the optimizations associated with directory-based caching algorithms. In this way, broadcasting overhead is restricted to a minimum to conserve bandwidth.

[0129] FIG. 26 illustrates a simplified data sharing and coherency process for a typical read operation to be performed on a resource in accordance with the present invention wherein the requested data is found on disk. First AMF 2300 (AMF#1) issues a shared data request 2310 to a second AMF 2350 (AMF #2) for the requested data. Accordingly, AMF #2 issues an acknowledgment request 2330 based on the shared data request 2310 issued from AMF 2300. However, in accordance with a preferred aspect of the invention, the shared data request is sent to N-1 replication partners 2350 which are sharing the resource with AMF 2300 (AMF #2 being one of such replication partners). If, the data is not found in the cache of a replication partner (cache miss), then a negative acknowledgment request 2370 is issued by the replication partner 2350. If the data is not found in the cache of a replication partner, the data is read directly from disk by AMF 2300 during operation step 2130, transferred to the host that requested the data, and the operation is completed in step 2140 (FIG. 23) in accordance with the read operation of the present invention.

[0130] While the invention has been described by way of example and in terms of the specific embodiments, it is to be understood that the invention is not limited to the disclosed embodiments. To the contrary, it is intended to cover various modifications and similar arrangements as would be apparent to those skilled in the art. Therefore, the scope of the appended claims should be accorded the broadest interpretation so as to encompass all such modifications and similar arrangements.

Claims

1. In a data storage network, a method of maintaining data coherency using two or more array management functions (AMFs) that are able to concurrently access a redundancy group, the redundancy group including a plurality of resources, the method comprising: receiving a request from a host, by a first one of the AMFs, to perform a first operation on data stored on a first one of the resources; broadcasting a message from the first AMF to the other AMFs sharing access to the first resource so as to acquire access to the first resource; and performing the first operation on the data by the first AMF.

2. The method of claim 1, wherein the first operation is a write operation, the method further comprising receiving from the host the data to be written to the first resource by the first AMF.

3. The method of claim 2, wherein the broadcast message is a write invalidate request, and wherein, responsive to the invalidate request, each of the AMFs sharing access to the resource invalidates corresponding data stored in its cache.

4. The method of claim 2, further comprising sending replication and state data from the first AMF to one or more other AMFs concurrently with performing the write operation.

5. The method of claim 1, wherein the first operation is a read operation.

6. The method of claim 5, further comprising performing a search of the cache of each of the AMFs sharing access to the first resource for a copy of the data requested in the read request.

7. The method of claim 6, further comprising reading the requested data from the first resource if none of the AMFs sharing access respond with a copy of the requested data.

8. The method of claim 6, further comprising receiving the requested data from one of the AMFs sharing access to the resource.

9. The method of claim 5, wherein the broadcast message identifies the data in the request, the method further comprising searching the cache of each of the AMFs sharing access to the first resource for the identified data, and forwarding the identified data to the first AMF if found in the cache.

10. The method of claim 1, further comprising determining whether the data identified in the request is shared by one or more other AMFs.

11. In a data storage network, a method of maintaining data coherency using two or more array management functions (AMFs) that are able to concurrently access a redundancy group, the redundancy group including a plurality of resources, the method comprising: receiving a request from a host, by a first one of the AMFs, to perform a first operation on data stored on a first one of the resources; determining from the request whether the identified data is shared by one or more of the other AMFs; and if so: broadcasting a message from the first AMF to the other AMFs sharing access to the first resource so as to acquire access to the first resource; and performing the first operation on the data by the first AMF; and if not: performing the first operation on the data by the first AMF.

12. The method of claim 11, wherein the request is one of a write data request and a read data request.

13. The method of claim 11, further comprising sending replication and state data from the first AMF to one or more other AMFs concurrently with performing the first operation.

14. The method of claim 11, wherein the broadcast message is a write invalidate request, and wherein, responsive to the invalidate request, each of the AMFs sharing access to the resource invalidates corresponding data stored in its cache.

15. A data storage network, comprising: two or more controllers communicably coupled to a redundancy group, each controller having one or more array management functions (AMFs), wherein two or more of said AMFs are able to concurrently access the redundancy group, the redundancy group including a plurality of resources, each controller further including: means for receiving a request from a host to perform an operation on data stored on a first one of the resources by a first AMF; means for broadcasting a message from the first AMF to other AMFs sharing access to the first resource so as to acquire access to the first resource; and means for performing the first operation on the data by the first AMF.

16. A data storage network, comprising: two or more controllers communicably coupled to a redundancy group, each controller having one or more array management functions (AMFs), wherein two or more of said AMFs are able to concurrently access the redundancy group, the redundancy group including a plurality of resources, each controller further including: means for receiving a request from a host to perform a first operation on data stored on a first one of the resources by a first AMF; means for determining from the request whether the identified data is shared by one or more of the other AMFs; means for broadcasting a message from the first AMF to the other AMFs sharing access to the first resource so as to acquire access to the first resource if the data is shared; and means for performing the first operation on the data by the first AMF.