Patent Grant
7797358
Computer data is vital to today's organizations, and a significant part of protection against disasters is focused on data protection. As solid-state memory has advanced to the point where cost of memory has become a relatively insignificant factor, organizations can afford to operate with systems that store and process terabytes of data.
Conventional data protection systems include tape backup drives, for storing organizational production site data on a periodic basis. Such systems suffer from several drawbacks. First, they require a system shutdown during backup, since the data being backed up cannot be used during the backup operation. Second, they limit the points in time to which the production site can recover. For example, if data is backed up on a daily basis, there may be several hours of lost data in the event of a disaster. Third, the data recovery process itself takes a long time.
Another conventional data protection system uses data replication, by creating a copy of the organization's production site data on a secondary backup storage system, and updating the backup with changes. The backup storage system may be situated in the same physical location as the production storage system, or in a physically remote location. Data replication systems generally operate either at the application level, at the file system level, or at the data block level.
Current data protection systems try to provide continuous data protection, which enable the organization to roll back to any specified point in time within a recent history. Continuous data protection systems aim to satisfy two conflicting objectives, as best as possible; namely, (i) minimize the down time, in which the organization production site data is unavailable, during a recovery, and (ii) enable recovery as close as possible to any specified point in time within a recent history.
One issue that may arise in CDP systems is that limited storage may be available for journal data. This may limit journal storage to less than is desirable.
The present invention provides methods and apparatus for a continuous data protection system having journal compression. With this arrangement, the amount of journal storage is decreased. While exemplary embodiments are shown and described in conjunction with particular configurations, it is understood that the inventive embodiments are applicable to systems in general in which it is desired to reduce journal storage.
In one aspect of the invention, a method includes storing journal data including a do data stream and an undo data stream in a continuous data protection system, compressing data prior to entry in the do data stream, storing compression information in a do metadata stream for the do data stream entry, accessing the data for the entry in the do data stream, examining the do metadata stream for the entry, decompressing the do data stream entry and writing the decompressed data to storage.
The method can further include one or more of the following features: compressing and decompressing data in the undo data stream, compressing the data based upon criteria, the criteria includes file type, the criteria includes CPU usage level, and the criteria includes application.
In another aspect of the invention, a system comprises a continuous data protection system to store journal data including a do data stream and an undo data stream, the continuous data protection system including a compression/decompression module to compress data prior to entry in the do data stream, store compression information in a do metadata stream for the do data stream entry, access the data for the entry in the do data stream, examine the do metadata stream for the entry, and decompress the do data stream entry and write the decompressed data to storage.
In a further aspect of the invention, an article comprises computer readable instructions to enable a machine to perform: storing journal data including a do data stream and an undo data stream in a continuous data protection system, compressing data prior to entry in the do data stream, storing compression information in a do metadata stream for the do data stream entry, accessing the data for the entry in the do data stream, examining the do metadata stream for the entry, and decompressing the do data stream entry and writing the decompressed data to storage.
The present invention will be more fully understood and appreciated from the following detailed description, taken in conjunction with the drawings in which:
The following definitions are employed throughout the specification and claims.
BACKUP SITE—a facility where replicated production site data is stored; the backup site may be located in a remote site or at the same location as the production site;
DPA—a computer or a cluster of computers that serve as a data protection appliance, responsible for data protection services including inter alia data replication of a storage system, and journaling of I/O requests issued by a host computer to the storage system;
HOST—at least one computer or networks of computers that runs at least one data processing application that issues I/O requests to one or more storage systems; a host is an initiator with a SAN;
HOST DEVICE—an internal interface in a host, to a logical storage unit;
IMAGE—a copy of a logical storage unit at a specific point in time;
INITIATOR—a node in a SAN that issues I/O requests;
JOURNAL—a record of write transactions issued to a storage system; used to maintain a duplicate storage system, and to rollback the duplicate storage system to a previous point in time;
LOGICAL UNIT—a logical entity provided by a storage system for accessing data from the storage system;
LUN—a logical unit number for identifying a logical unit;
PHYSICAL STORAGE UNIT—a physical entity, such as a disk or an array of disks, for storing data in storage locations that can be accessed by address;
PRODUCTION SITE—a facility where one or more host computers run data processing applications that write data to a storage system and read data from the storage system;
SAN—a storage area network of nodes that send and receive I/O and other requests, each node in the network being an initiator or a target, or both an initiator and a target;
SOURCE SIDE—a transmitter of data within a data replication workflow, during normal operation a production site is the source side; and during data recovery a backup site is the source side;
STORAGE SYSTEM—a SAN entity that provides multiple logical units for access by multiple SAN initiators
TARGET—a node in a SAN that replies to I/O requests;
TARGET SIDE—a receiver of data within a data replication workflow; during normal operation a back site is the target side, and during data recovery a production site is the target side;
WAN—a wide area network that connects local networks and enables them to communicate with one another, such as the Internet.
Reference is now made to
During normal operations, the direction of replicate data flow goes from source side to target side. It is possible, however, for a user to reverse the direction of replicate data flow, in which case Site I starts to behave as a target backup site, and Site II starts to behave as a source production site. Such change of replication direction is referred to as a “failover”. A failover may be performed in the event of a disaster at the production site, or for other reasons. In some data architectures, Site I or Site II behaves as a production site for a portion of stored data, and behaves simultaneously as a backup site for another portion of stored data. In some data architectures, a portion of stored data is replicated to a backup site, and another portion is not.
The production site and the backup site may be remote from one another, or they may both be situated at a common site, local to one another. Local data protection has the advantage of minimizing data lag between target and source, and remote data protection has the advantage is being robust in the event that a disaster occurs at the source side.
The source and target sides communicate via a wide area network (WAN) 128, although other types of networks are also adaptable for use with the present invention.
In accordance with an embodiment of the present invention, each side of system 100 includes three major components coupled via a storage area network (SAN); namely, (i) a storage system, (ii) a host computer, and (iii) a data protection appliance (DPA). Specifically with reference to
Generally, a SAN includes one or more devices, referred to as “nodes”. A node in a SAN may be an “initiator” or a “target”, or both. An initiator node is a device that is able to initiate requests to one or more other devices; and a target node is a device that is able to reply to requests, such as SCSI commands, sent by an initiator node. A SAN may also include network switches, such as fiber channel switches. The communication links between each host computer and its corresponding storage system may be any appropriate medium suitable for data transfer, such as fiber communication channel links.
In an embodiment of the present invention, the host communicates with its corresponding storage system using small computer system interface (SCSI) commands.
System 100 includes source storage system 108 and target storage system 120. Each storage system includes physical storage units for storing data, such as disks or arrays of disks. Typically, storage systems 108 and 120 are target nodes. In order to enable initiators to send requests to storage system 108, storage system 108 exposes one or more logical units (LU) to which commands are issued. Thus, storage systems 108 and 120 are SAN entities that provide multiple logical units for access by multiple SAN initiators.
Logical units are a logical entity provided by a storage system, for accessing data stored in the storage system. A logical unit is identified by a unique logical unit number (LUN). In an embodiment of the present invention, storage system 108 exposes a logical unit 136, designated as LU A, and storage system 120 exposes a logical unit 156, designated as LU B.
In an embodiment of the present invention, LU B is used for replicating LU A. As such, LU B is generated as a copy of LU A. In one embodiment, LU B is configured so that its size is identical to the size of LU A. Thus for LU A, storage system 120 serves as a backup for source side storage system 108. Alternatively, as mentioned hereinabove, some logical units of storage system 120 may be used to back up logical units of storage system 108, and other logical units of storage system 120 may be used for other purposes. Moreover, in certain embodiments of the present invention, there is symmetric replication whereby some logical units of storage system 108 are used for replicating logical units of storage system 120, and other logical units of storage system 120 are used for replicating other logical units of storage system 108.
System 100 includes a source side host computer 104 and a target side host computer 116. A host computer may be one computer, or a plurality of computers, or a network of distributed computers, each computer may include inter alia a conventional CPU, volatile and non-volatile memory, a data bus, an I/O interface, a display interface and a network interface. Generally a host computer runs at least one data processing application, such as a database application and an e-mail server.
Generally, an operating system of a host computer creates a host device for each logical unit exposed by a storage system in the host computer SAN. A host device is a logical entity in a host computer, through which a host computer may access a logical unit. In an embodiment of the present invention, host device 104 identifies LU A and generates a corresponding host device 140, designated as Device A, through which it can access LU A. Similarly, host computer 116 identifies LU B and generates a corresponding device 160, designated as Device B.
In an embodiment of the present invention, in the course of continuous operation, host computer 104 is a SAN initiator that issues I/O requests (write/read operations) through host device 140 to LU A using, for example, SCSI commands. Such requests are generally transmitted to LU A with an address that includes a specific device identifier, an offset within the device, and a data size. Offsets are generally aligned to 512 byte blocks. The average size of a write operation issued by host computer 104 may be, for example, 10 kilobytes (KB); i.e., 20 blocks. For an I/O rate of 50 megabytes (MB) per second, this corresponds to approximately 5,000 write transactions per second.
System 100 includes two data protection appliances, a source side DPA 112 and a target side DPA 124. A DPA performs various data protection services, such as data replication of a storage system, and journaling of I/O requests issued by a host computer to source side storage system data. As explained in detail hereinbelow, when acting as a target side DPA, a DPA may also enable rollback of data to an earlier point in time, and processing of rolled back data at the target site. Each DPA 112 and 124 is a computer that includes inter alia one or more conventional CPUs and internal memory.
For additional safety precaution, each DPA is a cluster of such computers. Use of a cluster ensures that if a DPA computer is down, then the DPA functionality switches over to another computer. The DPA computers within a DPA cluster communicate with one another using at least one communication link suitable for data transfer via fiber channel or IP based protocols, or such other transfer protocol. One computer from the DPA cluster serves as the DPA leader. The DPA cluster leader coordinates between the computers in the cluster, and may also perform other tasks that require coordination between the computers, such as load balancing.
In the architecture illustrated in
In accordance with an embodiment of the present invention, DPAs 112 and 124 are configured to act as initiators in the SAN; i.e., they can issue I/O requests using, for example, SCSI commands, to access logical units on their respective storage systems. DPA 112 and DPA 124 are also configured with the necessary functionality to act as targets; i.e., to reply to I/O requests, such as SCSI commands, issued by other initiators in the SAN, including inter alia their respective host computers 104 and 116. Being target nodes, DPA 112 and DPA 124 may dynamically expose or remove one or more logical units.
As described hereinabove, Site I and Site II may each behave simultaneously as a production site and a backup site for different logical units. As such, DPA 112 and DPA 124 may each behave as a source DPA for some logical units, and as a target DPA for other logical units, at the same time.
In accordance with an embodiment of the present invention, host computer 104 and host computer 116 include protection agents 144 and 164, respectively. Protection agents 144 and 164 intercept SCSI commands issued by their respective host computers, via host devices to logical units that are accessible to the host computers. In accordance with an embodiment of the present invention, a data protection agent may act on an intercepted SCSI commands issued to a logical unit, in one of the following ways:
Send the SCSI commands to its intended logical unit.
Redirect the SCSI command to another logical unit.
Split the SCSI command by sending it first to the respective DPA. After the DPA returns an acknowledgement, send the SCSI command to its intended logical unit.
Fail a SCSI command by returning an error return code.
Delay a SCSI command by not returning an acknowledgement to the respective host computer.
A protection agent may handle different SCSI commands, differently, according to the type of the command. For example, a SCSI command inquiring about the size of a certain logical unit may be sent directly to that logical unit, while a SCSI write command may be split and sent first to a DPA associated with the agent. A protection agent may also change its behavior for handling SCSI commands, for example as a result of an instruction received from the DPA.
Specifically, the behavior of a protection agent for a certain host device generally corresponds to the behavior of its associated DPA with respect to the logical unit of the host device. When a DPA behaves as a source site DPA for a certain logical unit, then during normal course of operation, the associated protection agent splits I/O requests issued by a host computer to the host device corresponding to that logical unit. Similarly, when a DPA behaves as a target device for a certain logical unit, then during normal course of operation, the associated protection agent fails I/O requests issued by host computer to the host device corresponding to that logical unit.
Communication between protection agents and their respective DPAs may use any protocol suitable for data transfer within a SAN, such as fiber channel, or SCSI over fiber channel. The communication may be direct, or via a logical unit exposed by the DPA. In an embodiment of the present invention, protection agents communicate with their respective DPAs by sending SCSI commands over fiber channel.
In an embodiment of the present invention, protection agents 144 and 164 are drivers located in their respective host computers 104 and 116. Alternatively, a protection agent may also be located in a fiber channel switch, or in any other device situated in a data path between a host computer and a storage system.
What follows is a detailed description of system behavior under normal production mode, and under recovery mode.
In accordance with an embodiment of the present invention, in production mode DPA 112 acts as a source site DPA for LU A. Thus, protection agent 144 is configured to act as a source side protection agent; i.e., as a splitter for host device A. Specifically, protection agent 144 replicates SCSI I/O requests. A replicated SCSI I/O request is sent to DPA 112. After receiving an acknowledgement from DPA 124, protection agent 144 then sends the SCSI I/O request to LU A. Only after receiving a second acknowledgement from storage system 108 will host computer 104 initiate another I/O request.
When DPA 112 receives a replicated SCSI write request from data protection agent 144, DPA 112 transmits certain I/O information characterizing the write request, packaged as a “write transaction”, over WAN 128 to DPA 124 on the target side, for journaling and for incorporation within target storage system 120.
DPA 112 may send its write transactions to DPA 124 using a variety of modes of transmission, including inter alia (i) a synchronous mode, (ii) an asynchronous mode, and (iii) a snapshot mode. In synchronous mode, DPA 112 sends each write transaction to DPA 124, receives back an acknowledgement from DPA 124, and in turns sends an acknowledgement back to protection agent 144. Protection agent 144 waits until receipt of such acknowledgement before sending the SCSI write request to LU A.
In asynchronous mode, DPA 112 sends an acknowledgement to protection agent 144 upon receipt of each I/O request, before receiving an acknowledgement back from DPA 124.
In snapshot mode, DPA 112 receives several I/O requests and combines them into an aggregate “snapshot” of all write activity performed in the multiple I/O requests, and sends the snapshot to DPA 124, for journaling and for incorporation in target storage system 120. In snapshot mode DPA 112 also sends an acknowledgement to protection agent 144 upon receipt of each I/O request, before receiving an acknowledgement back from DPA 124.
For the sake of clarity, the ensuing discussion assumes that information is transmitted at write-by-write granularity.
While in production mode, DPA 124 receives replicated data of LU A from DPA 112, and performs journaling and writing to storage system 120. When applying write operations to storage system 120, DPA 124 acts as an initiator, and sends SCSI commands to LU B.
During a recovery mode, DPA 124 undoes the write transactions in the journal, so as to restore storage system 120 to the state it was at, at an earlier time.
As described hereinabove, in accordance with an embodiment of the present invention, LU B is used as a backup of LU A. As such, during normal production mode, while data written to LU A by host computer 104 is replicated from LU A to LU B, host computer 116 should not be sending I/O requests to LU B. To prevent such I/O requests from being sent, protection agent 164 acts as a target site protection agent for host Device B and fails I/O requests sent from host computer 116 to LU B through host Device B.
In accordance with an embodiment of the present invention, target storage system 120 exposes a logical unit 176, referred to as a “journal LU”, for maintaining a history of write transactions made to LU B, referred to as a “journal”. Alternatively, journal LU 176 may be striped over several logical units, or may reside within all of or a portion of another logical unit. DPA 124 includes a journal processor 180 for managing the journal.
Journal processor 180 functions generally to manage the journal entries of LU B. Specifically, journal processor 180 (i) enters write transactions received by DPA 124 from DPA 112 into the journal, by writing them into the journal LU, (ii) applies the journal transactions to LU B, and (iii) updates the journal entries in the journal LU with undo information and removes already-applied transactions from the journal. As described below, with reference to FIGS. 2 and 3A-3D, journal entries include four streams, two of which are written when write transaction are entered into the journal, and two of which are written when write transaction are applied and removed from the journal.
Reference is now made to
Write transaction 200 generally includes the following fields:
one or more identifiers;
a time stamp, which is the date & time at which the transaction was received by source side DPA 112;
a write size, which is the size of the data block;
a location in journal LU 176 where the data is entered;
a location in LU B where the data is to be written; and
the data itself.
Write transaction 200 is transmitted from source side DPA 112 to target side DPA 124. As shown in
In practice each of the four streams holds a plurality of write transaction data. As write transactions are received dynamically by target DPA 124, they are recorded at the end of the DO stream and the end of the DO METADATA stream, prior to committing the transaction. During transaction application, when the various write transactions are applied to LU B, prior to writing the new DO data into addresses within the storage system, the older data currently located in such addresses is recorded into the UNDO stream.
By recording old data, a journal entry can be used to “undo” a write transaction. To undo a transaction, old data is read from the UNDO stream in a reverse order, from the most recent data to the oldest data, for writing into addresses within LU B. Prior to writing the UNDO data into these addresses, the newer data residing in such addresses is recorded in the DO stream.
The journal LU is partitioned into segments with a pre-defined size, such as 1 MB segments, with each segment identified by a counter. The collection of such segments forms a segment pool for the four journaling streams described hereinabove. Each such stream is structured as an ordered list of segments, into which the stream data is written, and includes two pointers—a beginning pointer that points to the first segment in the list and an end pointer that points to the last segment in the list.
According to a write direction for each stream, write transaction data is appended to the stream either at the end, for a forward direction, or at the beginning, for a backward direction. As each write transaction is received by DPA 124, its size is checked to determine if it can fit within available segments. If not, then one or more segments are chosen from the segment pool and appended to the stream's ordered list of segments.
Thereafter the DO data is written into the DO stream, and the pointer to the appropriate first or last segment is updated. Freeing of segments in the ordered list is performed by simply changing the beginning or the end pointer. Freed segments are returned to the segment pool for re-use.
When a write transaction is received, journaling is thus advanced as indicated in TABLE I below.
Conversely, during a rollback to undo a write transaction, the above operations are reversed, as indicated in TABLE II below.
The following example, in conjunction with
Three write transactions are received, as indicated in TABLE III.
The following discussion describes four stages of journaling and data storage; namely,
Stage #1: Enter the three write transactions as journal entries in the journal LU.
Stage #2: Apply the first write transaction to LU B.
Stage #3: Apply the second write transaction to LU B.
Stage #4: Rollback the second write transaction, to recover data from an earlier point in time.
The write transaction with ID=1 is written to the first 15 blocks of Segment #1. The metadata corresponding to this transaction is written to the first block of Segment #2. The second write transaction with ID=2 is written to the last 5 blocks of Segment #1 and the first 15 blocks of Segment #3. The metadata corresponding to this transaction is written to the second block of Segment #2. The third write transaction with ID=3 is written to the last 5 blocks of Segment #3 and the first 15 blocks of Segment #4. The metadata corresponding to this transaction is written to the third block of Segment #2.
Thus at stage #1, the DO stream in memory includes a list of segments 1, 3, 4; and a beginning pointer to offset=0 in Segment #1 and an end pointer to offset=10 in Segment #4. The DO METADATA stream in memory includes a list of one segment, namely Segment #2; and a beginning pointer to offset=0 in Segment #2 and an end pointer to offset=3 in Segment #2. The UNDO stream and the UNDO METADATA stream are empty. The journal and the four streams at the end of stage #1 are illustrated in
At stage #2 the write transaction with ID-1 is applied to LU B. New data to be written is read from the journal LU at the offset and length indicated in the DO METADATA; namely, 15 blocks of data located in blocks 0-14 of journal volume Segment #1. Correspondingly, old data is read from LU B at the offset and length indicated in the UNDO METADATA; namely, 15 blocks of data located in blocks 57-71 of LU B. The old data is then written into the UNDO stream in the journal LU, and the associated metadata is written into the UNDO METADATA stream in the journal LU. Specifically, for this example, the UNDO data is written into the first 15 blocks of Segment #5, and the UNDO METADATA is written into the first block of Segment #6. The beginning pointer of the UNDO data stream is set to offset=0 in Segment #5, and the end pointer is set to offset=15 in Segment #5. Similarly, the beginning pointer of the UNDO METADATA stream is set to offset=0 on Segment #6, and the end pointer is set to offset=1 in Segment #6.
At this point, the new data that was read from blocks 0-14 of journal LU Segment #1 is written to blocks 57-71 of LU B. The beginning pointer for the DO stream is moved forward to block 15 of journal LU Segment #1, and the beginning pointer for the DO METADATA stream is moved forward to block 1 of journal LU Segment #2. The journal and the four streams at the end of stage #2 are illustrated in
At stage #3 the write transaction with ID=2 is applied to the storage system. As above, 20 blocks of new data are read from blocks 15-19 of journal LU Segment #1 and from blocks 0-14 of journal LU Segment #3. Similarly, 20 blocks of old data are read from blocks 87-106 of LU B. The old data is written to the UNDO stream in the last 5 blocks of journal LU Segment #5 and the first 15 blocks of journal LU Segment #7. The associated metadata is written to the UNDO METADATA stream in the second block of Segment #6. The list of segments in the UNDO stream includes Segment #5 and Segment #7. The end pointer of the UNDO stream is moved to block 15 of Segment #7, and the end pointed of the UNDO METADATA stream is moved to block 2 of Segment #6.
Finally, the new data from blocks 15-19 of journal LU Segment #1 and blocks 0-14 of journal LU Segment #3 is written into blocks 87-106 of LU B. The beginning pointer for the DO stream is moved forward to block 15 of journal volume Segment #3, and the beginning pointer for the DO METADATA stream is moved forward to block 2 of journal LU Segment #2. Segment #1 is freed from the DO stream, for recycling within the segment pool, and the list of segments for the DO stream is changed to Segment #3 and Segment #4. The journal and the four streams at the end of stage #3 are illustrated in
At stage #4 a rollback to time 10:00:00.00 is performed. Since the write transaction with ID=3 was not applied yet, the only write transaction to be undone is the write transaction with ID=2. The last entry is read from the UNDO METADATA stream, the location of the end of the UNDO METADATA stream being determined by its end pointer, i.e., the metadata before block 2 of journal LU Segment #6 is read, indicating two areas each of 20 blocks; namely, (a) the last 5 blocks of journal LU Segment #5 and the first 15 blocks of journal LU Segment #7, and (b) blocks 87-106 of LU B. Area (a) is part of the UNDO stream.
The 20 blocks of data from area (b) are read from LU B and written to the beginning of the DO stream. As the beginning pointer of the DO stream is set to offset=15 of journal LU Segment #3, 5 blocks are written at the end of Segment #3, and the remaining 15 blocks are written to Segment #8. The end pointer for the DO stream is set to block 15 of Segment #8. The list of segments for the DO stream is changed to Segment #3, Segment #4 and Segment #8. The metadata associated with the 20 blocks from area (b) is written to block 3 of Segment #2, and the end pointer of the DO METADATA stream is advanced to block 4 of Segment #2.
The 20 blocks of data in area (a) of the journal LU are then written to area (b) of the LU B. Finally, Segment #7 is freed for recycling in the segment pool, the UNDO stream ending pointer is moved back to Segment #5 of the journal LU, block 15, and the UNDO METADATA stream ending pointed is moved back to Segment #6 of the journal LU, block 1. The journal and the four streams at the end of stage #4 are illustrated in
Thus it may be appreciated that the journal is thus used to rollback LU B to the state that it was in at a previous point in time. The journal is also used to selectively access data from LU B at such previous point in time, without necessarily performing a rollback. Selective access is useful for correcting one or more files that are currently corrupt, or for simply accessing old data.
TABLE IV below summarizes the behavior of the special protection components of system 100 during production mode. Reference is also made to
Only steps with arrows connecting them in
It is also noted in
Recovery mode is generally triggered as a result of a disaster at the source side. The source side data may become corrupt, or may not exist at all. In such case, after recovery is completed at the backup site, a user may perform a failover operation by switching the roles of the production site and backup site. The original backup site becomes a current production site, and the original production site becomes a current backup site. Alternatively, recovery mode can be triggered without a failover, in order to access data from a previous point in time.
While in recovery mode, target site DPA 124 continues to receive new write transactions from DPA 112 and enter them at the ends of the DO and DO METADATA streams. However, unlike production mode behavior, DPA 124 stops applying journal entries received from DPA 112 to LU B. Instead, DPA 124 uses the UNDO stream of the journal to rollback LU B, as described hereinabove.
During recovery, after or possibly before rollback of LU B is complete, a user may wish to access data from the target site. To this end, protection agent 164 stops failing I/O requests issued by host computer 160 and begins redirecting them to DPA 124. The processing of data by host computer 160 during recovery mode is referred to as “target side processing (TSP)”.
To manage TSP write commands that are received by target side DPA 124, journal processor 180 uses two additional data streams, referred to as TSP DO and TSP METADATA streams. When a TSP write command is received by DPA 124, it is entered at the end of the TSP DO stream and the end of the TSP DO METADATA stream. Since TSP writes relate to the state of LU B after the rollback is complete, the TSP DO stream writes are only applied to LU B after rollback is complete. Journal processor 180 applies TSP writes to LU B in a way similar to the way it applies write transactions deceiver from DPA 112; namely, journal processor 180 maintains the undo information for each write applied to LU B, in TSP UNDO and TSP UNDO METADATA streams.
When TSP read commands are received by target site DPA 124, DPA 124 returns the data to be read by identifying locations of the read command, and finding the most recent TSP write command or commands that were applied at these locations. The data is searched for (i) first in the TSP DO stream, and (ii) then in the journal UNDO data that was not yet applied to LU B and (iii) finally, if the data was not found in (i) and (ii), then the data is taken from LU B itself. In order to perform such a search efficiently, DPA 124 generates and stores in its memory a virtual image of the UNDO METADATA storage locations by using an efficient data structure, such as a binary search tree.
After rollback is completed, the TSP writes that were performed during the rollback are applied to LU B, and DPA 124 begins applying TSP writes synchronously; i.e., TSP writes are applied to LU B when they are received by DPA 124, without keeping them in the TSP DO stream. As such, when a read command is received after rollback is complete, it is sent directly to LU B instead of being redirected through DPA 124.
TABLES V and VI below summarize the behavior of the special protection components of system 100 during recovery mode, before and after the rollback is complete, in accordance with an embodiment of the present invention. Reference is also made to
It is also noted in
Reference is now made to
The data at time (1) is a common image for LU A and LU B at the same point in time, TCURRENT. Similarly, the data at time (2) is a common image for LU A and LU B at time TOLD. Rolled back data at time (2) may be processed by TSP writes, while at the same time current data at time (1) is being processed by source side writes. As such, the data evolves along the path from time (2) to time (3) as it is processed by the target side, and along the path from time (2) to time (4) as it is processed by the source side. The data images at the source and target sides at time TNEW are thus different.
When the recovery process is completed, the user may (i) return to a normal production mode, or (ii) perform a failover by switching the replication direction. In case (i), LU B is rolled back to its state at time (2), and the write transactions along the path from (2) to (4) are applied to LU B, so as to bring LU B to the same image as LU A. Conversely, in case (ii), LU B is maintained at its state at time (3), and its data is copied from the target side to the source side so as to bring LU A to the same image as LU B.
It may be appreciated that after rolling back the UNDO data stream to LU B, the state of the target side storage is substantially identical to the state that LU A was in at an earlier point in time. However, after applying TSP writes, the state of LU B is then in a new state that is different from the earlier state of LU A. As such, in order to return to a normal production mode, and ensure that LU B is a copy of LU A, DPA 124 undoes the TSP writes that were written to LU B using the TSP undo stream, and then returns to its normal production mode and begins applying the data that was written into the DO stream. The DO stream includes all write transactions that were undone while LU B was rolled back. Additionally, the DO stream includes new journal entries that were received from DPA 112 while DPA was in recovery mode. Similarly, protection agent 164 returns to its production mode by beginning to fail I/O requests issued by host 116.
Alternatively, the user want to perform a failover; i.e., to make LU B in its current state a production LU and ensure that LU A is a copy of LU B. In this case the write transactions in the DO stream that correspond to a point in time subsequent to the recovered point in time are ignored. Additionally, the TSP writes that were applied to LU B during the recovery process are applied to LU A. Thereafter, the replication direction changes. Specifically, DPA 124 and protection agent 164 begin behaving in accordance with source site behavior, and DPA 112 and protection agent 144 begin behaving in accordance with target site behavior.
It may be appreciated that in order to provide failover capability, in which the roles of the production site and the backup site are switched, it is desirable that the source side has the necessary system components to function as a target side, and vice versa. Thus, in an embodiment of the present invention, the source side includes its own journal LU 184 and journal processor 188, as indicated with dotted lines in
Referring back to TABLE I, it may be appreciated that during normal data replication, for each write transaction received from a production site, there are five I/O operations performed at a backup site. Reference is now made to
In accordance with an embodiment of the present invention, the meta-data for each transaction is of a fixed size, typically 30 bytes. The raw data varies in size, typically averaging around 10 KB per transaction.
As write transactions performed at a production site vary in frequency, and as each write transaction at the production site normally requires five I/O transactions at the backup site, it may be appreciated that the size of the DO stream grows and shrinks accordingly. When the I/O rate is low, the beginning of the DO stream is close to the end of the DO stream. In such case, it is possible to keep all write transactions between the beginning and the end of the DO stream in memory, and there is no need to read the beginning of the DO stream for every new transaction received in the backup site. As such, step 2 may be skipped.
Reference is now made to
However, during the steps 3-5 distribution of the write transaction that was just received, it is possible that a new transaction arrives at the backup site. In order to keep pace with the arriving transaction, as each write transaction is entered into the ends of the DO and DO METADATA streams, the write transaction is written into the end of a queue in memory. In accordance with an embodiment of the present invention, the queue in memory is handled similar to the way the DO stream is handled; namely, each received write is appended to the end of the queue, and when a write transaction is distributed according to steps 3-5, a subsequent write transaction is taken from the beginning of the queue. Effectively, the queue corresponds to a cached DO stream.
The 4-stage journaling process is used until the queue in memory is full, at which point the normal 5-stage journal processing is resumed. Also in the event of a disaster, the normal 5-stage journal processing is resumed. In order to resume the 5-stage journal processing, it is important to identify the last write in the DO stream that was written. As such, even during the 4-stage journal processing, the pointers to the first and last write transactions in the DO stream are updated.
Conversely, when the I/O rate is high, in order to control the size of the DO stream and ensure that it does not overflow its disk allotment, the present invention switches from the normal 5-stage mode to a faster 3-stage mode whenever the DO stream reaches a large percentage of its maximum capacity, typically 80%. The present invention afterwards switches back from the faster 3-stage mode to the normal 5-stage mode whenever the DO stream is reduced to a smaller percentage of its maximum capacity, typically 75%.
The 3-stage mode eliminates steps 3 and 4 from the normal mode; namely, the steps that record the UNDO information. As such, rollback of the backup storage unit to its state at the times of those transactions processed with the 3-stage mode is not possible.
Reference is now made to
TABLE VII summarizes the relative pros and cons of each of the journaling processes described hereinabove.
One data replication strategy is the set of automated rules for controlling when a data replication system transitions between 5-stage, 4-stage and 3-stage journal processing. As mentioned hereinabove, transitions from 5-stage to 3-stage journaling, and from 3-stage back to 5-stage journaling, may be controlled based on the current size of the DO stream. Transitions from 5-stage to 4-stage journaling may be automated to occur when the beginning and end of the DO stream are close; and transitions from 4-stage back to 5-stage journaling may be automated to occur when the memory queue reaches its capacity.
Reference is now made to
It will be appreciated by those skilled in the art that using 4-stage journaling enables a data replication system to keep pace with higher I/O rates than can be handled when using 5-stage journaling. If the system is currently using 5-stage journaling and the I/O rate is higher than can be handled, a lag increases until the system necessarily transitions to the 3-stage journaling process. However, if the system can catch up with the lag, empty the DO stream and transition to a 4-stage journaling process, then the system can accommodate higher I/O rates before transitioning back to the 5-stage journaling process.
In this regard, it is noted that in general, if the system cannot keep pace with I/O rates using a 4-stage journaling process, then it most probably cannot keep pace using a 5-stage journaling process, and in either case the system would have to transition to a 3-stage journaling process. However, since the I/O rate changes continuously, a transition from 4-stage journaling to 5-stage journaling does not necessarily push the system to 3-stage journaling.
Reference is now made to
In another aspect of the invention, a continuous data protection system includes journal compression. Data can be compressed prior to entry into the journal, i.e., at the I/O level. With this arrangement, journal data is compressed while enabling relatively rapid retrieval of decompressed journal data.
It is understood that the compression/decompression module 1302 can be provided as a software running on the journal processor 188, a separate device, a module running on a further processor, and/or a combination of hardware and software to meet the needs of a particular application.
In general, the compression/decompression module 1302 compresses data at the I/O level prior to entry into the journal data structures. For example, a write requested by the source side host 104 will result in the write data being compressed by the compression/decompression module 1302 before being placed in the DO stream. The DO metadata stream contains compression information, such as compression algorithm, parameters, etc. The DO stream data is decompressed before being written to the LUN. As described above, new DO data is placed at the end of the DO stream and data to be written to storage is at the beginning of the DO stream. Thus, the DO stream data is compressed and then uncompressed so that uncompressed data is written to storage to provide a replica of the source side storage. Similarly, data from the storage device (LUN) is read and compressed prior to placement in the UNDO stream. Going back the other way in time would result in the UNDO data being decompressed and written back to storage.
The methods and apparatus of exemplary embodiments of the invention may take the form, at least partially, of program code (i.e., instructions) embodied in tangible media, such as disks (element 5 in
Since the journal data is compressed, the journal requires less space than a uncompressed data. While compressing journal data may require processing resources, point in time images can be obtained with relative efficiency. Since compression is performed at the I/O level, if portions of the data are contained in the journal, only the (compressed) data associated with an I/O operation need to be decompressed.
In a further aspect of the invention, the particular compression mechanism can be selected based upon various criteria. Exemplary criteria include application, e.g., MICROSOFT WORD, ORACLE DATABASE, etc, CPU usage level, file type, file size, user option, journal size, journal fullness, etc. It is understood that a wide variety of criteria can be used to meet the needs of a particular application without departing from exemplary embodiments of the invention.
In an exemplary embodiment, the compression algorithm is selected based upon CPU usage and file type. If the CPU usage is greater than a selected threshold, compression is not performed in order to prioritize throughput. The compression algorithm is selected based upon file type. Word processing documents are compressed using a first algorithm, image documents are compressed using a second algorithm, databases are compressed using a third algorithm, etc. Compression parameters can be selected based on the same or different criteria. For example, more ‘expensive’ levels of compression will be avoided during times of intense CPU usage. In one embodiment, Lempel ziv based compression is used for text files data, no compression for already compressed format, and data base compression algorithms, which understand data base table format, for data bases.
In step 1504, the compressed data is written to the end of the DO stream. When preparing to write data to the storage, the compressed data is read from the beginning of the DO stream in step 1506. In step 1508, the data is decompressed using information from the metadata DO stream, which contains compression information needed for decompression. The decompressed data is then written to storage. Data for the UNDO stream is treated in a similar manner.
By compressing/decompressing the journal data, the amount of storage required for the journal data is reduced. Point in time images are readily generated using the DO and UNDO streams. While the images may be at the block level and not the file level, since each IO is tracked when an image is virtually mounted (which can be a file system, data base, etc.), it is not necessary to decompress a significant amount of data in order to read the file/table since the system compresses on I/O operations. Thus, files can be searched in the image with relative efficiency without searching the entire image. With this arrangement, it is not necessary to decompress huge objects.
While the invention is shown and described in conjunction with a particular embodiment having an illustrative architecture having certain components in a given order, it is understood that other embodiments well within the scope of the invention are contemplated having more and fewer components, having different types of components, and being coupled in various arrangements. Such embodiments will be readily apparent to one of ordinary skill in the art. All documents cited herein are incorporated herein by reference.
| Number | Name | Date | Kind |
|---|---|---|---|
| 5990899 | Whitten | Nov 1999 | A |
| 6621493 | Whitten | Sep 2003 | B1 |
| 6804676 | Bains, II | Oct 2004 | B1 |
| 7546485 | Cochran et al. | Jun 2009 | B2 |
| 20070162513 | Lewin et al. | Jul 2007 | A1 |
| 20070266053 | Ahal et al. | Nov 2007 | A1 |
| 20080082592 | Ahal et al. | Apr 2008 | A1 |
| 20080082770 | Ahal et al. | Apr 2008 | A1 |
