WO2011101482A1

WO2011101482A1 - Read-other protocol for maintaining parity coherency in a write-back distributed redundancy data storage system

Info

Publication number: WO2011101482A1
Application number: PCT/EP2011/052572
Authority: WO
Inventors: James Lee Hafner; Tarun Thukar
Original assignee: International Business Machines Corporation; Ibm United Kingdom Limited
Priority date: 2010-02-22
Filing date: 2011-02-22
Publication date: 2011-08-25
Also published as: GB2490400B; DE112011100623T5; CN102696017A; CN102696017B; US20110208996A1; DE112011100623B4; US8103904B2; GB2490400A; GB201206650D0

Abstract

Data storage reliability is maintained in a write-back distributed data storage system including multiple nodes. Each node comprises a processor and an array of failure independent data storage devices. Information is stored as a set of stripes, each stripe including a collection of at least a data strip and associated parity strips, the stripes distributed across a primary data node and multiple corresponding parity nodes. A read- other parity update protocol maintains parity coherency. The primary data node for each data strip drives parity coherency with the corresponding parity nodes, independently of other data nodes, in keeping relevant parity strips for the primary data node coherent. A parity value is determined based on data other than a difference between new data and existing data. A new parity value is based on new data and dependent data, wherein with respect to one data value, dependent data comprises other data encoded in a corresponding parity value.

Description

READ-OTHER PROTOCOL FOR MAINTAINING PARITY COHERENCY IN A WRITE-BACK DISTRIBUTED REDUNDANCY DATA STORAGE SYSTEM

FIELD OF THE INVENTION

The invention relates generally to maintaining data integrity. In particular, the present invention relates to parity coherency in data storage.

BACKGROUND OF THE INVENTION

In information technology (IT) systems, often data is stored with redundancy to protect against component failures resulting in loss of data. Such data redundancy can be provided by simple data mirroring techniques or via erasure coding techniques. Erasure codes are the means by which storage systems are made reliable. In erasure coding, data redundancy is enabled by computing functions of user data such as parity (exclusive OR) or other more complex functions such as Reed-Solomon encoding. A Redundant Array of Inexpensive Disks (RAID) stripe configuration effectively groups capacity from all but one of the disk drives in a disk array and writes the parity (XOR) of that capacity on the remaining disk drive (or across multiple drives). When there is a failure, the data located on the failed drive is reconstructed using data from the remaining drives.

When data is updated by a host device, the redundancy data (parity), must also be updated atomically to maintain consistency of data and parity for data reconstruction or recovery as needed. In most cases, such updates can be time consuming, as they usually involve many storage device accesses. To mitigate this effect, a redundancy system may employ a writeback or "fast write" capability wherein one or more copies of new host write data (i.e., host data and one or more copies thereof) are stored/written in independent cache components of the system. The write is acknowledged as complete to the host and the parity updates are delayed to a more suitable time (e.g., at de-stage time of new write data).

In monolithic systems (e.g., a controller with two redundant processors where all the storage disks are accessible to both processors), atomic parity update can be more easily managed by one of the processors with a full knowledge of events during the process. Recovery from error or interruption is simplified. However, in a distributed redundancy data storage system including a collection of loosely coupled processing nodes that do not share the same disks, there are many more components, less shared knowledge and many more failure states and events. Consequently, achieving atomic parity update is more difficult. "Distributed" means that it is a collection of nodes. "Redundant" means that it must have erasure coding. In a write-thru system (without fast write), if a parity update fails prior to acknowledgement of the write to the host, then the write fails and recovery is driven by the host. However, with a distributed redundancy storage system employing^? write, the host data is committed by the distributed redundancy storage system and must be reliably available at any future time.

Consequently, the atomic parity update must be managed internally within the distributed redundancy storage system.

US6983396B2 discloses a bridged controller for reducing the overhead of cache coherency processing on each of a plurality of primary controllers and increasing the overall throughput of the system. The bridged controller interfaces with dual-active pair of host-side controllers and the backend disk drive buses. The third controller allows a doubling of the number of backend busses reduces the overhead of cache coherency processing on each primary host- side controller and doubles the overall throughput.

US5627990 discloses a method that reduces the unnecessary duplication of data in the lower Cached Storage Drawer (CSD) subsystem and the upper Storage Controller (SC) cache, thereby permitting a higher effective cache hit ratio that results in higher overall storage subsystem performance.

US6542960B1 discloses a system and method for updating parity based upon locking and unlocking of a storage stripe in a RAID implementation in which the stripe includes a parity block. The stripe is locked to prevent colliding I/O operations from being performed thereto while a current I/O operation is underway with respect to the stripe.

US6446220B1 discloses a system for updating data. A first processing unit receives a data update to a data block in a first storage device. Parity data for the data block is maintained in a second storage device. A parity group is comprised of the data block and the parity data. After determining that the first processing unit does not control access to the parity group including the data block to update, the first processing unit sends a message to a second processing unit controlling access to the parity group requesting control of access to the parity group. The first processing unit determines new parity data from the data update, the data at the data block in the first storage device, and the parity data in the second storage device. The first processing unit then writes the data update to the data block in the first storage device and the new parity data to the second storage device.

US20080109616A1 discloses provide a method, system, and computer program product for optimizing I/O operations performed by a storage server operating on behalf of multiple clients to access data on a plurality of storage devices (disks). Embodiments of the present invention eliminate the need for selected read operations to write new data to physical data blocks by zeroing the physical data blocks to which new data will be written. Additionally, the need for reading old parity to compute new parity is eliminated. Instead, new parity is computed from the data to be written without the need of old parity or the storage server sends a command to a disk that stores parity. A module implemented at the disk that stores parity executes the command without reading, by the storage server, old parity.

BRIEF SUMMARY

Data storage reliability is maintained in a write-back distributed data storage system including multiple nodes, each node comprising an array of failure independent data storage devices. In one embodiment, information is stored as a set of stripes, each stripe comprising a collection of data strips and associated parity strips, the stripes distributed across multiple corresponding data nodes and multiple corresponding parity nodes. For each stripe, a data node maintains the associated data strip holding a first copy of data, and a parity node maintains a parity strip holding a parity for the collection of data strips, wherein parity coherency is maintained.

A read-other (RO) parity update protocol is performed for maintaining parity coherency, wherein a primary data node drives parity coherency with its corresponding parity nodes, independently of other data nodes, in order to keep its relevant parity strips coherent.

Performing a read-other parity update protocol further includes determining parity value based on data other than a difference between new data and existing data, such that a new parity value is computed based on new data and dependent data, wherein with respect to one data value, dependent data comprises other data encoded in a corresponding parity value.

Another embodiment comprises a write-back distributed data storage system including multiple nodes, each node comprising and an array of failure independent data storage devices. The nodes are configured for storing information as a set of stripes, each stripe including a collection of at least a data strip and associated parity strips, the stripes distributed across a primary data node and multiple corresponding parity nodes. The primary data node for each data strip is configured for maintaining the data strip holding a first copy of data, and each parity node configured for maintaining a parity strip holding a parity for the data strips.

The primary data node is further configured for performing a read-other parity update protocol for maintaining parity coherency, including driving parity coherency with the corresponding parity nodes, independently of other data nodes, in keeping its relevant parity strips coherent. Performing a read-other parity update protocol further includes determining parity value based on data other than a difference between new data and existing data, such that a new parity value is computed based on new data and dependent data, wherein with respect to one data value, dependent data comprises other data encoded in a corresponding parity value. Another embodiment comprises a computer program product for maintaining data storage reliability in a write-back distributed data storage system including multiple nodes, each node comprising an array of failure independent data storage devices. The computer program product comprises a computer readable storage medium having computer usable program code embodied therewith, the computer usable program code comprising computer usable program code configured to store information as a set of stripes, each stripe including a collection of at least a data strip and associated parity strips, the stripes distributed across a primary data node and multiple corresponding parity nodes. The computer usable program code further comprises computer usable program code configured for the primary data node for each data strip to maintain the data strip holding a first copy of data, and each parity node maintain a parity strip holding a parity for the data strips.

The computer usable program code further comprises computer usable program code configured for performing a read-other parity update protocol for maintaining parity coherency, including the primary data node driving parity coherency with the corresponding parity nodes, independently of other data nodes, in keeping its relevant parity strips coherent. The computer usable program code further comprises computer usable program code configured for determining parity value based on data other than the difference between new data and existing data, such that a new parity value is computed based on new data and dependent data, wherein with respect to one data value, dependent data comprises other data encoded in a corresponding parity value.

BRIEF DESCRIPTION OF THE DRAWINGS

Preferred embodiments of the present invention will now be described, by way of example only, with reference to the following drawings in which:

FIG. 1 shows an example distributed redundancy storage system;

FIG. 2 shows a block diagram of a system for read-other (RO) protocol for maintaining parity coherency in a write-back distributed redundancy storage system;

FIG. 3 shows a flowchart of a RO protocol process for maintaining parity coherency in a write-back distributed redundancy data storage system;

FIG. 4 shows an example embodiment of data parity layout for a distributed data storage system implementing a RAID6 erasure code to which the RO protocol is applicable;

FIG. 5 shows an example event sequence in a RO protocol process for maintaining parity coherency in a write-back distributed storage system implementing a RAID6 erasure code;

FIG. 6 shows an example embodiment of the data parity layout for a distributed data storage system implementing the Weaver erasure code to which the RO protocol is applicable; and FIG. 7 shows an example event sequence in a RO protocol process for maintaining parity coherency in a write-back distributed storage system implementing the Weaver erasure code data.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

The following description is made for the purpose of illustrating the general principles of the invention and is not meant to limit the inventive concepts claimed herein. Further, particular features described herein can be used in combination with other described features in each of the various possible combinations and permutations. Unless otherwise specifically defined herein, all terms are to be given their broadest possible interpretation including meanings implied from the specification as well as meanings understood by those skilled in the art and/or as defined in dictionaries, treatises, etc. The description may disclose several preferred embodiments providing data redundancy in information technology (IT) systems. While the following description will be described in terms of such systems for clarity and placing the invention in context, it should be kept in mind that the teachings herein may have broad application to all types of systems, devices and applications.

Most distributed storage systems either perform only write-thru or coordinate parity computations at the host or client. Coordinating parity computations enables recovery and coordination in a manner similar to a write-thru system. The write-thru and coordinating parity computations have similar latency costs, as more work is performed before the write is fully complete to the application. As such, improvement in reducing the complexity of parity update directly improves the distributed storage system performance and reliability.

A read-other (RO) protocol for maintaining parity coherency in a write-back distributed redundancy data storage system is provided. In one embodiment, the RO protocol employs a data-driven RO parity update protocol with fast write to improve processing for parity updates spread across nodes of the system and yet meets client guarantees for superior reliability and system design simplicity. The following terminology is provided to help in understanding the description further below.

Node a computer and disk enclosure with a distributed redundancy module, processor, memory, disks, power supply, etc., in a data storage system.

primary disk a disk within a node that holds a first copy of host data.

primary data node a node that holds the primary disk along with other disks.

user data data written by a host or client to a storage system for possible later retrieval.

parity data redundancy calculated by a scheme to recover lost user data. strip a chunk (block) of user data or parity.

stripe a collection of data strips and parity strips that are connected by parity calculations (an example is shown in FIG. 1 described further below). parity node a node that holds one of the parity data strips for a collection of data strips in a stripe.

RO read-other algorithm for computing parity.

dependent data with respect to one data value (e.g., a data strip), the dependent data is the other data (e.g., other data strips) encoded in a corresponding parity value (e.g., a parity strip).

driver a primary data node with new data in its cache which drives or begins the parity update protocol.

User data is stored as collection of stripes, each stripe comprising a collection of data strips and associated parity strips, distributed across multiple corresponding data owner nodes and multiple corresponding parity owner nodes. For each stripe, data owner node maintains the associated data strip holding a first copy of data, and a parity owner node maintains a parity strip holding a parity for the collection of data strips, wherein parity coherency is

maintained. Preferably, each parity calculated based on a different parity calculation formula. In an example RAID stripe, n+1 nodes are required to store the data of n independent nodes. FIG. 1 shows a distributed redundancy storage system stripe configuration with an array of nodes (i.e., NO, N4), each node having multiple disk drives (i.e., HDDO, HDD4), effectively groups capacity from all but one of the nodes and writes the parity P (e.g., XOR) of that capacity on remaining nodes or across multiple nodes (as shown). For example, the data strips D₀, ..., D₃ in HDDO of nodes NO, N3, respectively, are for user data, and a parity strip P₀_3 in HDDO of node N4 is for corresponding parity data (e.g., XOR of D₀, D₃). The data strips D₀, ..., D₃ and parity strip P₀_3 form a stripe. Spreading the parity data across the nodes as shown allows spreading the task of reading the parity data over several nodes as opposed to just one node. Writing on a node in a stripe configuration requires that one or more new parities be calculated for this new data and the new parities be written over the old parities in each of the nodes holding a parity. In this example, with respect to one data value (e.g., a data strip D₀), the dependent data is the other data (e.g., other data strips Di, D₂, D₃) encoded in a parity value (e.g., parity strip P_0-3). There is no restriction that information on a disk be only data or only parity. Other embodiments may have disks within a node containing primary data for one stripe and parity data for another stripe.

According to the data-driven RO parity update protocol, each primary data node generally acts independently of other data nodes, even if the same stripe, in keeping its relevant parity strips coherent and correct. For example in FIG. 1, primary data node NO owns data strip Do on primary disk HDDO and drives nodes N4 for coherency and reliability of its data or parity with respect to the rest of the stripe formed by D₀, ..., D₃, P_0-3.

FIG. 2 shows a write-back distributed redundancy data storage system 10 that implements a data-driven RO parity update protocol, according to an embodiment of the invention. The system 10 is utilized for storing data from a host system 11 and/or a client system 12. The system 10 includes multiple nodes 13. Each node comprises a processor 14, memory 15, a data manager including a distributed redundancy module 16, and an array of storage devices such as hard disk drives 17 (e.g., HDDO, HDDn). In one implementation, the distributed redundancy module 16 comprises program code for execution by the processor. In another implementation, the distributed redundancy module comprises firmware. The distributed redundancy module 16 in each node implements the data-driven RO parity update protocol for maintaining parity coherency in the data storage system 10 including. Parity update is performed in the system 10 that uses fast write (write-back) behavior at a time subsequent to the completion of any fast write operations. The RO parity update computes new parity from data only (i.e., new data at the primary data node and/or parity- integrated data from dependent data nodes), without computing a difference (delta) between new data and existing data. Each parity node of the primary data node sends read request messages to the other data nodes for their dependent data, and computes a new parity from the dependent data it receives and the new data of the primary data node.

Generally, each node 13 that owns a primary disk for a strip is responsible for driving the coherency and reliability of its data or parity with respect to the rest of the stripe. This is different from a data storage system in which one node is responsible for coherency either on a global level or on a per stripe basis. According to the data-driven RO parity update protocol described herein, since each primary data node generally acts independently of other data nodes, in keeping its relevant parity strips coherent and correct, the overall explicit coordination between nodes of the same stripe is reduced, and coordination is performed by less explicit means, such as protocol serialization (described further below). For the example implementations described herein, the distributed redundancy storage system 10 (FIG. 2) is fault tolerant to two storage device failures (either disk or node).

Those skilled in the art recognize that generalization to higher fault tolerance is within the scope of the invention. Each data strip is encoded into two parity strips. This is extendible to more than two parity strips for higher fault tolerant systems.

The data-driven RO parity update protocol enables a data node 13 to maintain (either in memory or on disk) at least two copies of data for a strip (new fast write data and parity integrated data). The data-driven RO parity update protocol is data driven since it is initiated by a primary data node to its parity nodes. Each node 13 is a primary data node for some data and a parity node for some other data, wherein each such node must be capable of performing each of the functions of the protocol (driver, first parity, second parity, etc.). The parity update protocol follows a forwarded or chained messaging protocol between the participants (i.e., from the primary data node through each of its parity nodes in a specified order, with responses following the reverse order). The parity update protocol is

implemented in each of the nodes 13 in FIG. 2. Each node 13 serves a primary data node for certain data, and serves as a parity node for certain other data, each node being capable of performing each of the functions of the protocol (driver, first parity, second parity, etc.). A data owning node initiates a parity data update protocol on the distributed storage system 10 employs fast write capabilities to compute parity data that is used to recover lost data. The parity data update protocol includes: a. a message forwarding protocol.

b. two phases: a preparation/lock phase and a commit/unlock phase.

c. a RO sub-protocol for new parity computations at each parity node.

d. a specified parity node forwarding order for each of the primary data nodes.

The RO sub-protocol is for the computation of new parity requiring reduced resources (disk I/Os, in particular), especially when the number of "other" reads is strictly less than that required for an alternative using read-modify- write subprotocol (e.g., three for a two fault tolerant erasure code). An implementation of the RO sub-protocol is well-suited to erasure codes on short stripes (two examples are provided further below). The fast write data is mirrored to all parity nodes that encode copies of that data (if not, then a simple modification to the protocol achieves this mirroring in the first step).

FIG. 3 shows a flowchart of a process 20 implementing an embodiment of the RO parity data update protocol involving a primary data node holding a data strip and corresponding multiple (e.g., two) parity nodes holding parity strips for that data strips. The process 20 includes:

Block 21: In a preparation phase, said primary data node sends a prepare and lock request message to its "first parity node" (according to a specified ordering), wherein the first parity node forwards that request to a "second parity node". Block 22: Each parity node of the primary data node sends read request messages to the other data nodes for their dependent data, and computes a new parity from the dependent data it receives and the fast write data (i.e., new data) of the primary data node (a difference between new and existing data is not computed).

Block 23: When the second parity node has prepared its new parity, the second parity node sends an acknowledge (ACK) reply message back to the first parity node indicating "ready". When the first parity node receives the ACK from the second parity node and has also prepared its new parity, the first parity node sends an ACK to the primary data node indicating "ready". This ends the prepare phase.

Block 24: In a commit/unlock phase, the primary data node sends a commit message to the first parity node, which is forwarded to the second parity node. Block 25: The second parity node "commits" its new parity value, unlocks its parity from update by another primary data node, and sends an ACK back to the first parity node indicating "done". Committing means marking the working copy of the parity as the "parity integrated value" and discarding the previous old value. The "parity integrated value" of parity is the value of parity that contains (is computed from) the parity integrated versions of all the data that it represents. The first parity node commits its new parity, unlocks its parity from update by another primary data node, and sends an ACK back to the primary data node indicating "done". The primary data node, on receipt of the first parity node's ACK, unlocks its new data from update, marks its new data as parity integrated and invalidates its old copy. This completes the RO parity update protocol.

The message communication may be via a communication line, host, bus connecting the nodes, etc., implementing a message forwarding protocol. If a node, other than dependent data nodes, participating in a RO parity update protocol (that is, the primary data node, the first parity node or the second parity node), fails during the preparation phase, the lowest survivor node in the forward sequence initiates an abort sequence with the survivor nodes. If a node fails during the commit phase, the lowest survivor node in the forward sequence restarts the commit sequence with the survivor nodes. If two nodes fail during each phase, the lone survivor node either aborts or commits by itself. An abort sequence is a forwarding message chain, similar to the commit/lock message chain that causes each surviving node to roll back its state to that before the start of the parity update protocol. In particular, locks are removed and no new parity is committed and no new data is marked as parity integrated.

The specific forward message ordering of the parity nodes for each primary data node and the locking/unlocking steps are arranged so that if the protocol is initiated concurrently by two or more primary data nodes that share one or more parity nodes, the protocol is deadlock and starvation free. When a parity node, which has its parity locked during a RO parity update protocol from one primary data node, receives a first parity update protocol message from another primary data node, the parity node can either reject the second parity update request (by an "abort/try again later" response) or may hold the second message until the completion of its unlock step for the first parity update protocol, at which point it may proceed with the second parity update protocol.

If a dependent data node fails to respond to a request from a parity node for dependent data, the parity node may be unable to complete the computation of its working parity value. In this situation, the parity node does not acknowledge "ready" on its reply during the preparation phase. Instead, it replies with "failure" notification. Such failure notification is forwarded to the primary data node; in response the primary data node initiates an "abort" message chain (instead of the "commit" chain) which causes the protocol to roll back the system state before the start of the protocol, and discard all locks and all working data. This is equivalent to any abort sequence initiated by any node after a failure during the protocol (e.g., failure of a node, or failure to read data from disk, etc.).

Embodiments of the invention provides a data driven protocol as well as

locking/forwarding/recovery. The explicit payloads and erasure code computation properties may vary as anyone skilled in the art will appreciate.

A recovery protocol is driven by a successor driver (i.e., lowest survivor in the forward sequence). The primary data node is the initial driver. If the driver fails, then the next node in the forwarding chain assumes that role. The role of the driver is to complete the RO parity update protocol by forcing it to go forward via a commit phase or to force it to roll back via an explicit abort phase.

After a failure, the surviving parity node closest to (and including if it survives) the primary data node in the forwarding messaging order assumes the role of a recovery driver and completes or aborts the protocol with the remaining survivor parity nodes.

As noted, each primary data node is capable of maintaining (either in memory or on disk, or both) at least two versions of the data for a strip. The first copy is the fast write copy comprising new data that was written into the storage system by a host or client but has not yet been integrated into the parity strips. The second copy is the parity-integrated copy that has been incorporated into the mathematical relationships defining the parity values. Fast write versions transition to the parity integrated version at completion of the parity update protocol at all relevant parity nodes (the prior parity-integrated version can be discarded as well).

Similarly, each parity node maintains a version of its parity strip value that contains (per the mathematical relationship defining that parity value) the current parity- integrated value for each of the data strips computed into that parity strip. Each parity node also manages a working copy that is computed as an intermediate value during the parity update protocol. At the end of the parity update protocol (as viewed by each parity node), the parity node transitions the working copy to the parity- integrated value and discards the previous parity- integrated value. It may also discard any fast write data and any dependent data used in the computation at this time.

For each primary data node, there is a defined ordered relationship of its parity nodes in sequence. For example, if a primary data node has two parity nodes P0 and PI, then the ordered relationship may be P0 < PI, or PI < P0. As such, each primary data node has a first parity node and a second parity node. The defined ordered relationship may be the same for each, or some, of the primary data nodes (e.g., if the primary data nodes share the same parity nodes). The ordered relationship may be different for each primary data node (e.g., if each primary data node has a different pair of parity nodes from compared to another primary data node). This set of ordered relationships (one for each primary data node) is specified for the entire stripe so as to avoid a deadlock or starvation condition among concurrently initiated RO parity update protocols. The ordering is dependent on the specific erasure code used.

As noted, the parity update protocol involves a preparation/lock phase and a commit phase. In one implementation, the fast write data for a primary data node has been mirrored during the fast write process to the affected parity nodes (otherwise, during the first step below, the fast write data is sent as payload in the first messages). The preparation phase proceeds as follows. First, the primary data node locks its fast write data from further update. The primary data node then sends a prepare and lock ("Prepare&Lock") message to the first parity node without a payload (except as noted above). In the prepare phase, parity nodes request the dependent data from dependent data nodes and compute parity directly (rather than from delta and old parity).

Then, the first parity node takes three actions. First, it locks its parity-integrated value from update by a different primary data node (only one such protocol can run concurrently). Second, the first parity node forwards the "Prepare&Lock" message (without data payload) to the second parity node. Third, the first parity node sends read requests to the dependent data nodes (dependent with respect to the primary data node initiating this protocol), requesting their parity-integrated copy of their data. Upon receipt of the requested parity- integrated information the first parity node computes a working parity value as the mathematical combination of the fast write value of the primary data node data value and the parity- integrated dependent data values (without destroying or utilizing the current parity- integrated value). The combination is determined by a mathematical formula (e.g., Reed- Solomon encoding) by which the parity value is computed. These last two actions may be taken in parallel or in series.

Upon receipt of the message from the first parity node, the second parity node locks its parity-integrated value from further update by a different primary data node, sends read requests to its dependent data nodes and on receipt of replies from the dependent data nodes, the second parity node also computes its new working parity value by the appropriate mathematical combination of the primary data node data value and the parity integrated dependent data values (the dependent data nodes for each parity node may be the same or may be different). When complete, the second parity node sends a "ready" ACK reply to the first parity node. When the first parity node has completed its tasks and received the "ready" ACK reply, the first parity node forwards the "ready" ACK reply back to the primary data node. Receipt of the "ready" ACK reply at the primary data node completes the lock/preparation phase.

In a commit phase, the primary data node sends a "Commit" request message to the first parity node, which in turn forwards the request to the second parity node. Upon receipt of the request, the second parity node transitions its working parity value to its parity-integrated value and responds to the first parity node with a "done" ACK reply. The second parity node also unlocks its parity-integrated value from update. The first parity node, and on receipt of "done" ACK reply transitions its working parity value to the parity- integrated value and also unlocks it. The second parity node then sends a final "done" ACK reply to the primary data node. The primary data node transitions its fast write version of the data to its parity-integrated version and unlocks it (allowing it to accept new fast write data). This completes the commit/unlock phase and the parity update protocol. There are two classes of node failures: those that are directly involved in the RO parity update protocol (primary data node and its parity nodes) and those that are ancillary (the dependent data nodes). If ancillary nodes fail after returning requested data to the parity nodes, then no action need be taken. If the ancillary nodes fail prior to returning requested data, then the requesting parity node cannot complete its preparation phase, wherein the parity node returns an "Error" during the preparation phase (followed by an Abort phase by the driver node).

If one or two key participant parity nodes in the parity update protocol either fail completely or fail because it can no longer perform the function of primary data node or first or second parity node (failure of other nodes has no effect on the parity update protocol) during the parity update protocol, a successor driver node handles the recovery with the survivor nodes. During the preparation/lock ("Prepare&Lock") phase, a node failure causes a roll back to the initial state, wherein the successor driver nodes initiates a forwarded Abort sequence with the survivor nodes. For example, if the first parity node fails, then the primary data node is the successor driver node and sends an abort to the second parity node. If the second parity node fails, then the primary data node is again the successor driver node and sends an abort to the first parity node. If the primary data node fails, then the first parity node is the successor driver node and initiates an abort sequence with the second parity node. The successor driver node is the survivor closest to (and possibly including) the primary data node in the forwarding order. Similarly, during the commit phase, a node failure causes a roll forward of the protocol. The successor driver node (as defined above) restarts the commit sequence. If only one node survives, then it autonomously performs recovery (abort in the preparation/lock phase) or commit (in the commit phase). Those skilled in the art recognize that generalization to higher fault tolerance is within the scope of the invention.

After a node failure, it is possible that a surviving node (specially a down-stream node) will have not received the initial message starting the protocol. Hence it may have no context for the abort message it receives. This is acceptable as it indicates that the node is in the intended state. Similarly, if a node failure occurs during the lock phase, a second commit message may arrive at a down-stream node. This again is acceptable since the node will have already completed the commit during the lock phase and is in the desired state.

In terms of interaction between instances of parity update protocols, at the primary data node during the parity update protocol, no new fast write data can be integrated into the fast write version used to compute the delta value. There are alternative embodiments. First, additional buffer space can be utilized to hold other fast write data for the strip, in effect having a third buffer space for fast write data not involved in parity update protocol.

Second, the lock on the fast write version need only be set at the start of the lock phase. To accommodate this, if new fast write data arrives at the primary data node during the prepare/lock phase, and then instead of completing the protocol with the commit phase, the primary data node can abort the protocol with an explicit abort phase. This is similar to the commit phase in message ordering and replies, but the abort instruction informs the parity nodes to discard their working parity value and to not transition them to the parity-integrated value (the protocol rolls back to the initial state). If a parity node (first or second) receives a parity update request ("Prepare&Lock") from a second primary data node during the processing of a parity update protocol from a first primary data node, the parity node can either delay responding to the new request until the first parity update protocol is complete, or reject the second request indicating that a parity update protocol is in progress. The second primary data node can retry the protocol at a later time. The specific ordering of the parity nodes for each primary data node prevents starvation in the delay case and deadlock in the reject case. In either case, the first parity update protocol will be able to reach completion.

The parity update protocol may be extended to higher fault tolerance. If a data strip is encoded into three or more parity values on three or more parity nodes, then the parity update protocol again orders each of the parity nodes in a specific way. The protocol is then forwarded in each of the preparation/lock phase and the commit phases from the primary data node to each of the parity nodes according to the specific order for that primary data node. If one or more nodes fail during the protocols, the successor survivor node (defined by the node closest to, and possibly including, the primary data node) drives the completion of the protocol (i.e., abort during the preparation/lock phase and commit during the commit phase).

Two example implementations of read-modify- write for maintaining parity coherency are described below.

RAID6

FIG. 4 shows an example distributed redundancy data storage system 30, illustrating data and parity relations for the RAID6 erasure code on three data nodes (NO through N2) and two parity nodes N3, N4, with fault tolerance two. RAID6 provides a storage system with a striped set and dual distributed parity with fault tolerance from two disk drive failures (the storage system can continue operation when one or two disk drives fail). The dependent data nodes for each parity node are the same, as both P and Q are computed from DO, Dl and D2. The RO parity update protocol is useful with this RAID6 erasure code when the number of data nodes is two or three. FIG. 4 shows three data strips DO through D2. According to embodiments of the invention, a P parity is computed with a mathematical formula (e.g., exclusive OR) of the three data strips DO through D2. Further, a Q parity is computed for the three data strips DO through D2 by a different mathematical formula (e.g., Reed- Solomon encoding). Other mathematical formulas may also be used. In this erasure code there are three primary data nodes and two parity nodes. The forwarding order for each primary data node is the same: P < Q, meaning P comes before Q in the forwarding order (the alternate order of Q < P is also possible). There may be variations where the order is P < Q for some primary data nodes and Q < P for others. A common order (e.g., P < Q) is sufficient to prevent deadlock or starvation.

FIG. 5 shows an event sequence 40 (top to bottom) for the parity update protocol implemented by system 30 in FIG. 4 according to an embodiment of the invention, wherein primary data node NO initiates the parity update protocol. The expression {dO'} indicates fast write data. In this example, the fast write data is mirrored on three different nodes. The expression QOR represents a mathematical formula used to compute or update the Q-parity value. The nodes Nl and N2 are ancillary nodes providing dependent read data. The read requests for the dependent data comes independently from both parity nodes N3 and N4, and using caching only one disk read is needed at each of the nodes Nl and N2.

The expressions [dO], [dl], and [d2] represent data versions that are integrated into the parity represented by [pO-2] and [qO-2]. The expressions pO-2' represents the working copy of the new parity maintained on parity node P. Similarly, the expression qO-2' represents the working copy of the new parity maintained on parity node Q. DiskRead[dl] and

DiskRead[d2] each represent the operation of reading from disk the respective parity integrated data strip. The expressions Read[dl] and Read[d2] represent a message request from a parity node to a the respective dependent data node to send the parity integrated copy of the dependent data. The expressions Reply[dl] and Reply[d2] are the return messages with payload containing the respective dependent data. The expression Prep&Lock indicates a message between two nodes that requests a lock on the parity strip. The expressions lock[d0'], lock[p0-2], lock[q0-2], unlock[d0'], unlock[p0-2'] and unlock[q0-2'] represent the actions of locking and unlocking respectively, the indicated data or parity strip. The expressions pO-2'=XOR(dO',dl,d2) and qO-2'=QOR(dO',dl,d2) represent the mathematical relationship whereby pO-2' is computed from d0', dl and d2 and qO-2' is computed from d0', dl, and d2, respectively. The expression Ready indicates a (good) ACK to the

Prep&Lock request; the expression Done indicates a (good) ACK to the Commit request message.

An alternative implementation that reduces network traffic may be implemented as follows. Before parity node N3 forwards a "Prepare&Lock" request to parity node N4, parity node N3 first issues read requests to nodes Nl and N2 for dependent data. Parity node N3 then computes the new Q parity working value (on behalf of parity node N4), as well as its own working value. Then parity node N3 forwards the "Prepare&Lock" request message to parity node N4 with a payload containing the Q working parity value. Then parity node N4 performs a lock and responds to parity node N3 immediately with an ACK message indicating "Ready". The parity node N3 then forwards the ACK back to the primary data node NO. This results in less network traffic, but leaves the state at the end of the

Preparation phase the same as above.

FIG. 5 represents an event sequence diagram wherein each vertical line represents a node in the storage system that holds and is responsible for data in a RAID6 stripe with 3 data strips and parities P and Q ( RAID6(3+P+Q)). Other nodes in the storage system not related to this stripe are not represented. The vertical bars labeled NO, Nl, and N2 represent nodes that each hold one of the three data strips in the stripe. The vertical bars labeled P and Q, respectively, represent nodes that hold the P parity value and Q parity value for the stripe, respectively. Below the labels are one or two symbol expressions. The expression [dO] represents the parity integrated version of the data for strip 0 stored on node NO. Similarly, expressions [dl] and [d2], [pO-2], and [qO-2] represent the parity integrated versions of the data strips dl and d2 and the P parity strip pO-2 (computed from the parity integrated data d0,dl,d2), and the Q parity stripe qO-2 (also computed from the parity integrated data d0,dl,d2), respectively. The expression {dO'} under NO, P and Q represents new fast write data for strip 0 that has been replicated on the three nodes NO, P and Q and not yet integrated into the parity values. In FIG. 5, arrows represent communication messages sent from the node at the tail of the arrow to the node at the head of the arrow. Thick arrows represent messages that contain data (more generally, bulk payloads), and thin arrows represent control messages or responses that have no bulk payloads. The dashed line across the middle represents a transition between the two phases of the protocol. In the first phase ("Preparation Phase"), an interruption in the protocol is handled by a roll back of the protocol. In the second phase ("Commit Phase"), an interruption is handled by a roll forward of the protocol. The annotation at the bottom indicates the "forwarding rule" in embodiments of the invention for the RAID6 version of the protocol, as does the header "NO initiates, N0-> P-> Q",(i.e., a data node initiates the protocol by the first message to P, which forwards to Q (responses follow the reverse path)). Nodes Nl and N2 are ancillary nodes providing only read data to the protocol. Time is represented top-down in the FIG. 5. The relative size of the time gaps does not necessarily represent time to scale. They represent only relative ordering in time.

The protocol is initiated by the node NO and proceeds as follows. In the first phase

(Preparation), NO locks the {d0'} version of the strip 0 data ("lock[d0']") from further update (e.g., by a new host write). This is to guarantee that no change is made to the value d0' until the protocol completes the integration of the current d0' into the parity. A variant of the protocol can relax this requirement. It may mark {d0'} as "in preparation for parity integration". If a new write arrives during the preparation phase, then the protocol may be safely aborted (rolled back). The absolute locking of {d0'} is only required during the Commit Phase.

Once the lock[d0'] is in place, NO sends a message ("Prep&Lock") to the P parity node. The message indicates: (a) start a Preparation for parity integration of {d0'}, and (b) lock parity [pO-2] so that no other node can proceed with the same protocol. No bulk data is sent in this initial message. Node P then takes three actions: it locks [pO-2] as instructed, it forwards the message to the node Q and it sends read requests ("Read[dl]" and "Read[d2]") to the ancillary nodes Nl and N2 requesting a copy of their parity integrated versions of their strip data. On receipt of the "Prep&Lock" message from P, node Q performs the following actions analogous to P's. Node Q locks its [qO-2] from update and sends read request messages to the ancillary nodes Nl and N2 for their parity integrated data. The lock step at Q is not essential when the forwarding rule of the protocol sends the first message always through P since the lock at P will preclude any conflicting messages from arriving at Q (e.g., in FIG. 7 this is not the case).

When Q has received all of the requested read responses, it computes a new qO-2' from the new dO' and the received dl and d2, using the QOR function for the Q parity value. At this point, Q sends the "Ready" response to node P. Similarly, when P has received all of its requested read responses, it computes a new value pO-2' from the new dO' and the received dl and d2 using the simple XOR function. After completing this computation and receiving the "Ready" response from Q, node P sends a "Ready" response to the originating data node NO. On receipt of this response from P, NO completes its part in the Preparation Phase and the Preparation Phase is complete.

The Commit Phase begins with a message "Commit" from NO to P which forwards it to Q. On arrival at Q, Q commits (transitions) the computed copy qO-2' as the "integrated" copy, unlocks its parity value (enabling other updates to occur) and sends "Done" response message back to P. Q has now completed its role in the protocol. P, on receipt of the "Done" response from Q, performs the analogous actions: commits (transitions) the computed copy pO-2' as the "integrated" copy, unlocks its parity value (enabling other parity update protocols from other nodes) and responds to NO with "Done" message. Finally, NO, on receipt of this "Done" response from P, commits its dO' copy as the "integrated copy" and unlocks dO', allowing new updates to be written to the same address location. This completes the Commit Phase (it also concludes the protocol).

Two variants of this protocol may be used to reduce the network bandwidth requirements in the Preparation Phase. In the first variant, P forwards the "Prep&Lock" message to Q and Q locks its parity qO-2 and responds with "Lock-In-Place". Q does not issue any read commands for data from Nl and N2. When P receives the read data from Nl and N2, P computes both pO-2' and qO-2' (on behalf of Q). Then P sends the computed value qO-2' to Q who responds with "Ready" message. In the second variant, P does not forward the "Prep&Lock" request directly. Instead, it waits for the read data from Nl and N2, computes the values of pO-2' and qO-2' (again on behalf of Q). P then appends the qO-2' data to the "Prep&Lock" message it now forwards to Q. Q responds with "Ready" message as before (without needing to send read requests to Nl and N2). At this point, the state of P and Q is identical as to the state in FIG. 5 immediately before the sending of P's "Ready" response to NO. The rest of the protocol completes as described above. These two variants reduce the network bandwidth requirements: only one bulk message is sent to Q (the computed value of qO-2') rather than two bulk messages ([dl] and [d2]). For wider RAID6 stripes, there is still only one bulk message to Q in these variants, but additional read response messages from other ancillary nodes.

WEA VER

FIG. 6 shows an example distributed redundancy data storage system 50, illustrating data stripes DO through D3 and parity (P23, P03, P01, PI 2) for the Weaver code on 4 nodes (NO through N3), with fault tolerance two. The design features of Weaver codes include: (a) placement of data and parity blocks on the same strip, (b) constrained parity in-degree, and (c) balance and symmetry. These codes are in general not maximum distance separable (MDS) but have optimal storage efficiency among all codes with constrained parity in- degree. Weaver codes are described in James Lee Hafner, "WEAVER Codes: Highly Fault Tolerant Erasure Codes for Storage Systems," published at

http://www.usenix.org/events/fast05/teclVfullj3apers/hamer_weaver/hafner_weaver.pdf December 2005. This erasure code is useful with the RO parity update protocol as each parity value is computed as the XOR (exclusive OR) of only two data values.

In FIG. 6, each node performs three roles in this erasure code. Each node is a primary data node and each node is also a parity node for two other primary data nodes. In this erasure code, each data node has two parity nodes that it needs to coordinate with. The two parity nodes are different for each primary data node. The ordering of first and second parity node for each primary data node is performed lexicographically with respect to the parity held by the nodes. Accordingly, N2 < Nl < N3 < NO since NO holds the parity P01, Nl holds P03, N3 holds P12 and NO holds P23. Each data node sends its first parity update protocol message to the lowest parity node in this order that has parity for that data node, and the lowest parity node forwards the protocol message to the next lowest parity node that has parity for that data. Specifically, the following message traffic ensues:

NO sends message to N2 which forwards to Nl,

Nl sends message to N2 which forwards to N3,

N2 sends message to N3 which forwards to NO,

N3 sends message to Nl which forwards to NO.

The dependent data nodes for each parity node are different. P01 is computed from DO and Dl, and P03 is computed from DO and D3. As such, for P01 (on node N2), the dependent data relative to DO is Dl and for P03 (on node Nl), the dependent data for DO is D3. That is, each parity node sees a different piece of dependent data for DO.

FIG. 7 shows an event sequence 60 (top to bottom) for the parity update protocol implemented by system 50 of FIG. 6, according to an embodiment of the invention, wherein node NO initiates the protocol. The expression {d0'} indicates fast write data. In this example, the fast write data is mirrored on three different nodes.

FIG.7 represents an embodiment of the invention for the WEAVER code (two fault tolerant on 4 nodes). The structure of FIG. 7 is similar to that of FIG. 5. The vertical bars represent nodes in the system, however, in this case there are only 4 nodes, labeled NO, Nl, N2 and N3. Time ordering is represented top-down, arrows indicate messages, with double arrows indicating bulk messages (with payload) and single arrows indicating control messages (without bulk payload). For this code, each node has a responsibility for both a data node function and a parity node function. In analogy with FIG. 5, N2 plays the role of P and N2 plays the role of Q for the data node NO. N2 and Nl hold parity values (pOl and p03, respectively) into which the data dO is computed. The forwarding rule for NO is N0->N2- >N1. As indicated at the bottom of FIG. 7, the forwarding rule is different for each data node (this is in contrast to FIG. 5, RAID6 where the order is always the same for each data node). For example, node Nl sends its parity update protocol messages first to N2 and then to N3. One embodiment of the rule is defined by a lexicographic ordering of the nodes by the parity labels. Other embodiments are possible so long as there are no cycles. Below the node labels are symbols that represent parity integrated versions of data and parity. For example, on node NO, [d0,p23] indicates that NO is responsible as the primary owner for data on stripO and also the parity owner for the parity p23=XOR(d2,d3). The brackets ([]) indicate that these two values are the parity integrated values. Similarly, the expressions [dl,p03] for Nl, [d2,p01] for N2 and [d3,pl2] for N3 represent the parity integrated values of dl,d2 and d3 and the parity values p03=XOR(d0,d3), p01=XOR(d0,dl) and pl2=XOR(dl,d2). The expressions {d0'} below the labels for NO, Nl and N2 indicate that each of these nodes is holding a new version of data for strip 0 (d0') that is not yet parity integrated.

The remainder of the description of the sequence diagram in FIG. 7 is similar to that for FIG. 5 on replacing P by N2 and Q by Nl, replacing pO-2 with pOl and qO-2 with p03, and replacing QOR with a simple XOR. There are only two significant differences: (a) there is only one ancillary node read request required for each parity node, and (b) some messages can be merged for efficiency. This is detailed in the following description, for further clarity.

As in FIG. 5, node NO initiates the protocol to update its parity by first locking its copy {d0'} from update. It then sends the "Prep&Lock" request to its first parity node N2

(holding pOl). Node N2 locks its parity from update and forwards the "Prep&Lock" message to the second parity node in the NO's forward order (Nl holding p03). To facilitate better efficiency, N2 includes in the "Prep&Lock" message a read request for the ancillary data it needs from Nl, namely, "Read[dl]", a read request for the parity integrated version Nl 's data dl (this may be performed by separate messages). On receipt of this message, node Nl takes three actions (the order is not important): (a) it locks its parity p03 from update, (b) it performs a disk read ("DiskRead[dl]"), and (c) it sends a read request

("Read[d3]") to its ancillary node N3. Node N3 reads the data [d3] from the disk and sends the data in response back to the requestor Nl . On receipt, Nl computes its new parity value p03' from the new data d0' and the received data d3. Node Nl then sends a response to node N2 to the initial "Prep&Lock+Read[dl]" message indicating both that it is "Ready"

(has completed preparation of p03') and including the bulk data dl requested in the Read portion of the message (if sent as separate messages, this dual response may be performed as separate responses). After computing its new parity value pOl ' from the new data d0' and the received data [dl], node N2 sends the "Ready" response back to the initial requestor NO. This completes the Preparation Phase of the protocol. The Commit Phase of the protocol is identical to that described above for FIG.5 on replacing P by N2, Q by Nl, pO-2' with pOl ' and qO-2' with p03'.

The non-cyclic ordering in FIG. 7 meets the conditions to prevent deadlock or starvation (enabling at least one of two concurrent parity update protocols to proceed, whether or not the other is rejected or paused).

Referring back to FIGs. 4 and 6, in contrast to the RAID6 example in FIG. 4, for the Weaver implementation in FIG. 6, the nodes Nl and N2 (holding parity for data in node NO) have different dependent data requirements. In addition, node N2 requires dependent data of node Nl (node Nl requires dependent data of node N3). Further, in FIG. 7, a read request for dependent data from node N2 to node Nl is merged with the "Prepare&Lock" request. This is not a requirement, but it does minimize messaging overhead at the cost of serializing the read requests for dependent data. In an alternative implementation, node N2 can forward only the "Prepare&Lock" request to node Nl . Then, each of node Nl and node N2 may concurrently issue their own read requests for dependent data. When node Nl receives the read response and computes its working parity value, it can respond to the "Prepare&Lock" request message from node N2 with an ACK message indicating "Ready". When node N2 has received both the "Ready" response and the dependent read response from node Nl, and has also computed its working parity value, then node N2 may respond to node NO with an ACK message indicating "Ready", completing the preparation phase of the RO parity update protocol.

In the expressions [d0,p23], [dl,p03], [d2,p01], and [d3,pl2], the subexpressions dO, dl,d2 and d3 represent data versions integrated into their respective parity versions. The subexpressions p23, p03, pOl, and p 12 are parity the integrated parity for the data with the indicated number label (so pOl is the integrated parity for dO and dl). The expressions p03' represents the working copy of the new parity maintained on node Nl which is the primary data node for dl and the second parity node for dO and the first parity node for d3. Similarly, the expression ρΟΓ represents the working copy of the new parity maintained on node N2 (the primary data node for d2 and the first parity node for both dO and for dl). The expressions Read[dl] and Read[d3] represent a message request from a parity node to a the respective dependent data node to send the parity integrated copy of the dependent data. The expressions Reply[dl] and Reply[d3] are the return messages with payload containing the respective dependent data. DiskRead[dl], and DiskRead[d3] each represent the operation of reading from disk the respective parity integrated data strip. The expression Prep&Lock indicates a message between two nodes that requests a lock on the related data or parity strip.

The expressions lock[d0'], lock[p03], lock[p01], unlock[d0'], unlock[p03'] and

unlock[p01 '] represent the actions of locking and unlocking respectively, the indicated data or parity strip. The expressions p03'=XOR(d0', d3) and pOl '=XOR(d0',dl) represent the mathematical relationship whereby p03 ' is computed from dO' and d3 and pOl ' is computed from dO' and dl, respectively. The arrow diagram at the bottom of FIG.7 indicates the forwarding order for each primary data node to its first and second parity node. The expression Ready indicates a (good) ACK to the Prep&Lock request; the expression Done indicates a (good) ACK to the Commit request message.

As such, a data-driven RO protocol for parity update on a distributed redundancy storage system with fast write provides the following features. The parity ordering for each primary data node prevents two RO parity update protocols that overlap at a parity node from running concurrently while allowing for forward progress to at least one of parity protocols. In one example, an initiating primary data node initially locks out parity update protocols for three nodes (itself, and two others). A first message from the initiating node is directed to a node in the protocol chain of the remaining nodes. The first set of messages (i.e., the first set of forwarding messages of the prepare and lock phase (not the ACK replies)) to reach an intersecting node dominate, while the other messages are delayed or rejected in the first phase (i.e., prepare and lock phase). Two RO parity update protocols that do not overlap at either parity node may run simultaneously (concurrently). In the presence of a failure during a RO parity update protocol, in the preparation phase, the RO parity update protocol rolls back wherein the successor node in the forward order expressly aborts the protocol during this phase (all work is discarded and retried when all the nodes are stabilized). Further, in the commit phase, the protocol will generally roll forward wherein the successor node in the forward order re-drives the commit protocol to ensure consistency among the survivor nodes. The RO parity update protocol is applicable to any two or more fault tolerant erasure codes, provided the parity ordering for each primary data node is defined appropriately. As noted, since in the RO parity update protocol there are two types of nodes (i.e., active participant nodes and ancillary nodes), then failure of an ancillary node during operation of a corresponding RO protocol leads to immediate roll-back. This is because the RO protocol cannot complete due to unavailable data from an ancillary node. Failure of an active participant nodes results in either roll back (during the preparation phase) or roll forward (during the commit phase).

The RO parity update protocol does not require interaction or knowledge or recovery of state to any node that failed and was replaced during the protocol. The survivor nodes of the original protocol are the only nodes in the completion (either roll back or roll forward) of the protocol.

The number of fast write copies of new data that are maintained in the storage system, or where they are maintained, is not essential since the copies need not be held at the parity nodes. If one or more fast write copies of new data are held within one or more of the parity nodes for that data, such copies may be discarded once the parity update protocol is complete at a parity node. If one or more fast write copies of new date are held elsewhere in the storage system, then an explicit "discardable" message from the primary data node to the fast write holder enables release of the resources used for those copies. If the fast write copies of new data are not held at every parity node, then a Prepare&Lock message from any node holding that copy may contain the new data within its payload, thereby providing that copy to a parity node which does not hold the new fast write data. As is known to those skilled in the art, the aforementioned example architectures described above, according to a preferred embodiment of the present invention, can be implemented in many ways, such as program instructions for execution by a processor, as software modules, microcode, as computer program product on computer readable media, as logic circuits, as application specific integrated circuits, as firmware, etc. The embodiments of the invention can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment containing both hardware and software elements. In a preferred

embodiment, the invention is implemented in software, which includes but is not limited to firmware, resident software, microcode, etc.

Furthermore, the embodiments of the invention can take the form of a computer program product accessible from a computer-usable or computer-readable medium providing program code for use by or in connection with a computer, processing device, or any instruction execution system. For the purposes of this description, a computer-usable or computer readable medium can be any apparatus that can contain, store, communicate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. The medium can be electronic, magnetic, optical, or a semiconductor system (or apparatus or device). Examples of a computer-readable medium include, but are not limited to, a semiconductor or solid state memory, magnetic tape, a removable computer diskette, a RAM, a read-only memory (ROM), a rigid magnetic disk, an optical disk, etc. Current examples of optical disks include compact disk-read-only memory (CD-ROM), compact disk-read/write (CD-R/W) and DVD.

I/O devices (including but not limited to keyboards, displays, pointing devices, etc.) can be connected to the system either directly or through intervening controllers. Network adapters may also be connected to the system to enable the data processing system to become connected to other data processing systems or remote printers or storage devices through intervening private or public networks. Modems, cable modem and Ethernet cards are just a few of the currently available types of network adapters. In the description above, numerous specific details are set forth. However, it is understood that embodiments of the invention may be practiced without these specific details. For example, well-known equivalent components and elements may be substituted in place of those described herein, and similarly, well-known equivalent techniques may be substituted in place of the particular techniques disclosed. In other instances, well-known structures and techniques have not been shown in detail to avoid obscuring the understanding of this description. The terms "computer program medium," "computer usable medium," "computer readable medium," and "computer program product," are used to generally refer to media such as main memory, secondary memory, removable storage drive, a hard disk installed in hard disk drive, and signals. These computer program products are means for providing software to the computer system. The computer readable medium allows the computer system to read data, instructions, messages or message packets, and other computer readable information, from the computer readable medium. The computer readable medium, for example, may include non- volatile memory, such as a floppy disk, ROM, flash memory, disk drive memory, a CD-ROM, and other permanent storage. It is useful, for example, for

transporting information, such as data and computer instructions, between computer systems.

Furthermore, the computer readable medium may comprise computer readable information in a transitory state medium such as a network link and/or a network interface, including a wired network or a wireless network that allow a computer to read such computer readable information. Computer programs (also called computer control logic) are stored in main memory and/or secondary memory. Computer programs may also be received via a communications interface. Such computer programs, when executed, enable the computer system to perform the features of the present invention as discussed herein. In particular, the computer programs, when executed, enable the processor or multi-core processor to perform the features of the computer system. Accordingly, such computer programs represent controllers of the computer system.

Generally, the term "computer-readable medium", as used herein, refers to any medium that participated in providing instructions to a processor for execution. Such a medium may take many forms, including but not limited to, non- volatile media, volatile media and

transmission media. Non-volatile media includes, for example, optical or magnetic disks. Volatile media includes dynamic memory. Transmission media includes coaxial cables, copper wire and fiber optics, including the wires that comprise a bus. Transmission media can also take the form of acoustic or light waves, such as those generated during radio wave and infrared data communications. Common forms of computer-readable media include, for example, a floppy disk, a flexible disk, hard disk, magnetic tape, or any other magnetic medium, a CD-ROM, any other optical medium, punch cards, paper tape, any other physical medium with patterns of holes, a RAM, a PROM, an EPROM, a FLASH-EPROM, any other memory chip or cartridge, a carrier wave as described hereinafter, or any other medium from which a computer can read.

The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems that perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.

The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and/or

"comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof.

The corresponding structures, materials, acts, and equivalents of all means or step plus function elements in the claims below are intended to include any structure, material, or act for performing the function in combination with other claimed elements as specifically claimed. The description of the present invention has been presented for purposes of illustration and description, but is not intended to be exhaustive or limited to the invention in the form disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope of the invention. The embodiment was chosen and described in order to best explain the principles of the invention and the practical application, and to enable others of ordinary skill in the art to understand the invention for various embodiments with various modifications as are suited to the particular use contemplated.

While certain exemplary embodiments have been described and shown in the accompanying drawings, it is to be understood that such embodiments are merely illustrative of and not restrictive on the broad invention, and that this invention not be limited to the specific constructions and arrangements shown and described, since various other modifications may occur to those ordinarily skilled in the art.

Claims

1. A method of maintaining data storage reliability in a write-back distributed data storage system including multiple nodes, each node comprising an array of failure independent data storage devices, the method comprising:

storing information as a set of stripes, each stripe including a collection of at least a data strip and associated parity strips, the stripes distributed across a primary data node and multiple corresponding parity nodes;

the primary data node for each data strip maintaining the data strip holding a first copy of data, and each parity node maintaining a parity strip holding a parity for the data strips; and

performing a read-other parity update protocol for maintaining parity coherency, including the primary data node driving parity coherency with the corresponding parity nodes, independently of other data nodes, in keeping relevant parity strips for the primary data node coherent; and

wherein performing a read-other parity update protocol further includes determining parity value based on data other than a difference between new data and existing data, such that a new parity value is computed based on new data and dependent data, wherein with respect to one data value, dependent data comprises other data encoded in a corresponding parity value.

2. The method of claim 1, wherein:

each parity is calculated based on a different parity calculation formula; and performing a read-other parity update protocol further comprises determining parity by:

in a preparation phase:

the primary data node sending a prepare message to a first parity node, each parity node forwarding the prepare message to a next parity node in a predefined sequence;

each parity node sending request messages to the other data nodes for their dependent data, and computing a new parity from received dependent data and at least one of the new data at the primary data node and parity-integrated data from dependent data nodes; starting from the last parity node in the sequence, each parity node sending an acknowledgement message to an immediately prior parity node in the sequence; and

the first parity node sending an acknowledgement message to the primary data node, indicating completion of the preparation phase.

3. The method of claim 2, wherein performing a read-other parity update protocol further comprises:

in a commit phase:

the primary data node sending a commit message to the first parity node, each parity node forwarding the commit message to a next parity node in a predefined sequence;

each parity node committing its newly computed parity value;

starting from the last parity node in the sequence, each parity node sending an acknowledgement message to an immediately prior parity node in the sequence; and

the first parity node sending an acknowledgement message to the primary data node, indicating completion of the commit phase.

4. The method of claim 3 further comprising:

the primary data node maintaining at least two versions of the data for a strip, a fast write version and a parity- integrated version, wherein the fast write copy has been written into a write-back distributed data storage system by a host or client and not integrated into a parity strip, and wherein the parity-integrated copy has been incorporated into a process for defining a parity value; and

transitioning the fast write version to the parity- integrated version at completion of a parity update protocol at all relevant parity nodes, whereby the prior parity- integrated version is discarded.

5. The method of claim 4 further comprising, in each parity node:

maintaining a working version of its parity strip value containing a current parity- integrated value for each of the data strips computed into that parity strip; and

at the end of the parity update as viewed by each parity node, the parity node transitioning the working version to the parity- integrated value and discarding the previous parity-integrated value.

6. The method of claim 1 further comprising:

during the preparation phase, upon a node failure, the lowest survivor node in the forward sequence driving an abort sequence for the read-other parity update protocol; and during the commit phase, upon a node failure, the lowest survivor node in the forward sequence driving a restart of the commit phase for the read-other parity update protocol.

7. The method of claim 1 further comprising:

storing information further comprises storing information as a set of stripes, each stripe including a collection of multiple data strips and associated parity strips, the stripes distributed across multiple corresponding primary data nodes and multiple corresponding parity nodes;

the primary data node for each data strip maintaining the data strip holding a first copy of data, and each parity node maintaining a parity strip holding a parity for the multiple data strips;

storing information as plural stripes, each stripe including a collection of a data strip and associated parity strips, distributed across a primary data node and multiple

corresponding parity nodes;

providing a specified parity node message forwarding order for each primary data node and corresponding parity nodes; and

each primary data node driving parity coherency with its corresponding parity nodes, independently of other data nodes, in keeping relevant parity strips for the primary data node coherent.

8 A write-back distributed data storage system for maintaining data storage reliability, comprising:

multiple nodes, each node comprising an array of failure independent data storage devices;

the nodes configured for storing information as a set of stripes, each stripe including a collection of at least a data strip and associated parity strips, the stripes distributed across a primary data node and multiple corresponding parity nodes, the primary data node for each data strip configured for maintaining the data strip holding a first copy of data, and each parity node configured for maintaining a parity strip holding a parity for the data strips;

the primary data node further configured for performing a read-other parity update protocol for maintaining parity coherency, including driving parity coherency with the corresponding parity nodes, independently of other data nodes, in keeping relevant parity strips for the primary data node coherent; and

9. The system of claim 8, wherein:

each parity is calculated based on a different parity calculation formula; and in a preparation phase :

the primary data node sends a prepare message to a first parity node, each parity node forwarding the prepare message to a next parity node in a predefined sequence;

each parity node sends request messages to the other data nodes for their dependent data, and computes a new parity from received dependent data and at least one of the new data at the primary data node and parity-integrated data from dependent data nodes;

starting from the last parity node in the sequence, each parity node sends an acknowledgement message to an immediately prior parity node in the sequence; and

the first parity node sends an acknowledgement message to the primary data node, indicating completion of the preparation phase.

10. The system of claim 9 wherein:

in a commit phase:

the primary data node sends a commit message to the first parity node, each parity node forwards the commit message to a next parity node in a predefined sequence;

each parity node commits its newly computed parity value;

starting from the last parity node in the sequence, each parity node sends an acknowledgement message to an immediately prior parity node in the sequence; and the first parity node sends an acknowledgement message to the primary data node, indicating completion of the commit phase.

11. The system of claim 10, wherein:

the primary data node maintains at least two versions of the data for a strip, a fast write version and a parity- integrated version, wherein the fast write copy has been written into a write-back distributed data storage system by a host or client and not integrated into a parity strip, and wherein the parity-integrated copy has been incorporated into a process for defining a parity value.

12. The system of claim 11, wherein each parity node is further configured for:

13. The system of claim 8, wherein:

during the preparation phase, upon a node failure, the lowest survivor node in the forward sequence drives an abort sequence for the read-other parity update protocol; and during the commit phase, upon a node failure, the lowest survivor node in the forward sequence drives a restart of the commit phase for the read-other parity update protocol.

14. The system of claim 8, wherein:

the nodes are configured for storing information as plural stripes, each stripe including a collection of a data strip and associated parity strips, the stripes distributed across a primary data node and multiple corresponding parity nodes;

a specified parity node message forwarding order is maintained for each primary data node and corresponding parity nodes; and

each primary data node is configured for driving parity coherency with its corresponding parity nodes, independently of other data nodes, in keeping relevant parity strips for the primary data node coherent.

15. A computer program product for maintaining data storage reliability in a writeback distributed data storage system including multiple nodes, each node comprising an array of failure independent data storage devices, the computer program product comprising a computer readable storage medium having computer usable program code embodied therewith, the computer usable program code comprising:

computer usable program code configured to store information as a set of stripes, each stripe including a collection of at least a data strip and associated parity strips, the stripes distributed across a primary data node and multiple corresponding parity nodes; computer usable program code configured for the primary data node for each data strip to maintain the data strip holding a first copy of data, and each parity node maintain a parity strip holding a parity for the data strips; and

computer usable program code configured for performing a read-other parity update protocol for maintaining parity coherency, including the primary data node driving parity coherency with the corresponding parity nodes, independently of other data nodes, in keeping relevant parity strips for the primary data node coherent, wherein performing a read other parity update protocol further includes determining parity value based on data other than a difference between new data and existing data, such that a new parity value is computed based on new data and dependent data, wherein with respect to one data value, dependent data comprises other data encoded in a corresponding parity value.

16. The computer program product of claim 15 further comprising:

computer usable program code configured for calculating each parity based on a different parity calculation formula, in a preparation phase:

each parity node sending request messages to the other data nodes for their dependent data, and computing a new parity from received dependent data and at least one of the new data at the primary data node and parity-integrated data from dependent data nodes;

starting from the last parity node in the sequence, each parity node sending an acknowledgement message to an immediately prior parity node in the sequence; and the first parity node sending an acknowledgement message to the primary data node, indicating completion of the preparation phase.

17. The computer program product of claim 16 further comprising:

computer usable program code configured for, in a commit phase:

each parity node committing its newly computed parity value;

18. The computer program product of claim 17 further comprising:

computer usable program code configured for the primary data node to maintain at least two versions of the data for a strip, a fast write version and a parity-integrated version, wherein the fast write copy has been written into a write-back distributed data storage system by a host or client and not integrated into a parity strip, and wherein the parity- integrated copy has been incorporated into a process for defining a parity value;

computer usable program code configured for transitioning the fast write version to the parity- integrated version at completion of a parity update protocol at all relevant parity nodes, whereby the prior parity-integrated version is discarded;

computer usable program code configured for each parity node to:

maintain a working version of its parity strip value containing a current parity- integrated value for each of the data strips computed into that parity strip; and

at the end of the parity update as viewed by each parity node, transition the working version to the parity- integrated value and discarding the previous parity- integrated value.

19. The computer program product of claim 15 further comprising: computer usable program code configured for, during the preparation phase, upon a node failure, the lowest survivor node in the forward sequence driving an abort sequence for the read-other parity update protocol; and

computer usable program code configured for, during the commit phase, upon a node failure, the lowest survivor node in the forward sequence driving a restart of the commit phase for the read-other parity update protocol.

20. The computer program product of claim 15 further comprising:

computer usable program code configured for storing information as plural stripes, each stripe including a collection of a data strip and associated parity strips, the stripes distributed across a primary data node and multiple corresponding parity nodes;

computer usable program code configured for providing a specified parity node message forwarding order for each primary data node and corresponding parity nodes; and computer usable program code configured for each primary data node driving parity coherency with its corresponding parity nodes, independently of other data nodes, in keeping relevant parity strips for the primary data node coherent.