Patent Grant
6954877
The present invention relates generally to computing systems, and more particularly to a method and apparatus for checkpointing in computing systems. Checkpointing allows processes to save their state periodically so that they can be restarted in the event of failure.
In many computing environments, it is desirable to provide continuous operation, even in the event of component failure. To maintain operation of a computing system during a component failure requires fault tolerance. One known technique for achieving fault tolerance employs a redundant process pair. The primary process actually performs the work and periodically synchronizes a backup process with the primary process using checkpointing techniques. With prior known checkpointing techniques, the primary sends messages that contain information about changes in the state of the primary process to the backup process. Immediately after each checkpoint, the primary and backup processes are in the same state.
In other prior known checkpointing methods, distinctions between operations that change state (such as write operations) and operations that do not change the state (such as read operations) are not made, and all operations are checkpointed to the backup process. Such a system is shown in U.S. Pat. No. 4,590,554 (Glazer—Parallel Computer Systems) where all inputs to the primary are provided via messages and all messages sent to the primary are made available to the secondary or backup, essentially allowing the backup to “listen in on” the primary's messages. Another such system is described in and U.S. Pat. No. 5,363,503 (Gleeson—Unisys Corporation) where checkpointing is provided as described in U.S. Pat. No. 4,590,554.
Other prior art, such as that shown in U.S. Pat. No. 4,228,496 (Katzman—Tandem Computers), describe that the primary receives a message, processes the message, and produces data. The produced data is stored in the primary's data space thereby changing the primary's data space. The change in the primary's data space causes a checkpointing operation of the data space to be made available to the backup. Thus, there is frequent copying of the primary's data space to the backup's data space, which uses a significant amount of time and memory for transferring the state of the primary to the backup. It may also result in the interruption of service upon failure of the primary. The overhead for such checkpointing methods can have considerable performance penalties.
Other prior art examples attempt to update only portions of the state of the primary that has changed since the previous update, but use complex memory and data management schemes. In others as shown in U.S. Pat. No. 5,621,885 (Del Vigna—Tandem Computers) the primary and backup, which run on top of a fault tolerant runtime support layer (that is, an interface between the application program and operating system) are resident in memory and accessible by both the primary and backup CPUs used in the described fault-tolerance model. The primary and backup processes perform the same calculations because they include the same code.
[Possible other prior art: Pat. Nos. 5,455,932; 5,157,663; 4,823,256; 5,155,678; 5,968,185; 5,802,265]
In addition, systems that provide fault tolerance and implement conventional checkpointing schemes are physically equivalent in their implementation or at least share resources.
In light of the above, it is desirable to arrive at an approach to checkpointing that may be used to foster fault-tolerance without some or all of the drawbacks to conventional checkpointing approaches described above.
According to the disclosed embodiments, a computing system provides a mechanism for checkpointing in a fault-tolerant service. The service is made fault tolerant by using a process pair; the primary process performs the work officially, while one or more backup processes provide a logical equivalent that can be used in the event of failure.
In accordance with an aspect of the invention, it provides checkpointing that allows the primary and backup to be logically equivalent at any given point in time, but may be internally different physically or in their implementation.
According to another aspect of the present invention, it provides checkpointing for a primary and backup that do not share physical resources, their failure domains are independent, and they do not have any version conflicts with respect to the internal state of their private resources.
Another aspect of the present invention provides checkpointing for a primary and backup that communicate with each other via network protocols.
Another aspect of the present invention provides lightweight or simplified checkpointing for a primary and backup by checkpointing only external requests or messages that changes the state of the service instance. Checkpointing is not applied to read operations.
Another aspect of the present invention provides checkpointing that allows rolling upgrades to new versions of software or equipment.
In view of the foregoing aspects and others, the present invention provides a system and method for logical checkpointing. According to the disclosed embodiments, checkpointing is lightweight and greatly simplified as a result of checkpointing only the external view. This is in contrast to conventional approaches that provide checkpointing of the internal state.
FIGS. 2A1 to 2A3 are block diagrams of software configurations according to an embodiment.
FIGS. 2B1 to 2B3 are block diagrams of hardware configurations according to an embodiment.
The present invention will now be described in detail with reference to the drawings, which are provided as illustrative examples of the invention so as to enable those skilled in the art to practice the invention. Notably, the implementation of certain elements of the present invention may be accomplished using software, hardware or any combination thereof, as would be apparent to those of ordinary skill in the art, and the figures and examples below are not meant to limit the scope of the present invention. Moreover, where certain elements of the present invention can be partially or fully implemented using known components, only those portions of such known components that are necessary for an understanding of the present invention will be described, and detailed descriptions of other portions of such known components will be omitted so as not to obscure the invention. Further, the present invention encompasses present and future known equivalents to the known components referred to herein by way of illustration. The preferred embodiment of the present invention will be described herein with reference to an exemplary implementation of a fault-tolerant service in a distributed computing environment. However, the present invention is not limited to this exemplary implementation, but can be practiced in any computing environment including a distributed storage system.
Looking now to
A distributed, scalable, and fault-tolerant service 140, which is a component of the highly scalable computing system 100, can be made up of one or more primary and backup service instance pairs. A primary service instance (106 and 108) is made up of a primary server (102a and 102n respectively) and can include a persistence mechanism including but not limited to a database (103a and 103n respectively) while a backup service instance (107 and 109) is made up of a backup server (104a and 104n respectively) and can include a database (105a and 105n respectively). The database portions are not essential to an embodiment, but can be part of the service instances. The terms primary server and primary (used by itself) can be used interchangeably. Similarly, the terms backup server and backup (used by itself) can be used interchangeably. It should be noted that a primary service instance could have more than one backup service instance associated with it. The primary service instance and its one or more backup service instances reside on separate failure domains, for example, but not limited to physical machines, to ensure independent failure, thereby avoiding the primary service instance and its one or more backup service instances failing together.
In accordance with the principles of the present invention, these components are provided for purposes of illustration only and are not limiting. The services can be any service, for example but not limited to, a highly scalable service and a fault-tolerant service. Of course, while only two services are shown 130 and 140, many more of each of these services and servers may be connected to one another via a network. In the example shown, the highly scalable service 130 includes a plurality of clients, while service 140 includes a fault-tolerant process pair.
At any point in time, the primary (102a-102n) and backup servers (104a-104n) are logically equivalent but may be physically, or in their implementation, entirely different internally. This means that even during the process of migrating the server pair to different resources such as a new version of software or new hardware equipment, the primary and the one or more backups have the same logical internal state. If the servers start processing from the same initial state and the same requests are applied in parallel on both sides (on the primary and backup), their internal logical states, notwithstanding the difference in the underlying hardware and/or software resources, are the same. It should be noted that the protocol used to migrate to new equipment or software, that is, whether availability is required or not, can be determined by customer needs. An embodiment can support both protocols, that is, it can support making the system continuously available or temporarily unavailable during these transitions. In this way, an embodiment allows migration to different resources, both software and hardware, to be achieved very easily and with or without disruption.
In accordance with the embodiments, the results of processing incoming requests are the same not because the primary and backup servers have identical internal data, but rather because they execute the same requests in parallel. The primary and backup servers can be running different versions of software and continue to have logically the same data and provide the same responses to requests. In other words, processed inputs are the same at the primary and the backup, the processing may or may not be the same, but the results of the processing are the same at both the backup and the primary.
An embodiment of the present invention can also accommodate complete software migration, which can include rolling upgrades (the process of upgrading software without making a service unavailable). See FIG. 2A3 for an example of a configuration showing completed software migration. To achieve complete software migration without disruption, bring down the second backup (104a), which is currently running Version 1 software (step 213). Use the boot server to bring the server (104a) up again with Version 2 software and checkpoint data to it and bring server (104a) current (step 214). As soon as the backup is current, switch over a backup (104b or 104a) and use it as a primary running Version 2 (step 215). The switchover process can be automated by the software or can be done manually. Server (102a) can remain off line and maintained as a spare or it can be rebooted with Version 2 software and placed back into service (step 216). The system now has server 104b running Version 2 software as primary, server 104a running Version 2 software as backup, and depending on whether 102a is maintained as a spare or made operational, perhaps 102a running Version 2 software as a backup (step 217). Software migration is complete without the system becoming unavailable during the transition to new software.
A corollary to software migration is hardware migration. If the new backups are created on different machines, then over time, migration to different machines allows replacement of older equipment.
If complete hardware migration is desirable, bring up 107c as a backup service on Model B equipment and checkpoint data to it (step 224). As soon as backup service 107c is brought current, remove services 106 and 107, which are running on Model A equipment (step 225). Removing 106 causes one of the remaining services 107b or 107c, which is running on Model B equipment to become the primary service. The selection of the machine that is to be the primary can be done automatically or manually. In an embodiment of the present invention, machine selection is typically done automatically. The primary service receives and processes all requests from the clients and the old primary 106 running on Model A equipment can be removed without disruption. Hardware migration is completed and the primary (107b) and backup 107c are running on Model B equipment (step 226). FIG. 2B3 shows a completed hardware migration configuration. The steps for partial and complete hardware migration can vary from the steps listed in FIG. 2B. This is an example for illustration purposes only and the order in which services are removed and brought up can vary to achieve migration.
Paramount to migration to different resources, that is, to different hardware or software, is the share-nothing nature of the service instances, that is, service instances do not share any physical resources. Their failure domains are independent and they do not have any version conflicts with respect to the internal state of their private resources. All resources, for example, but without limitation disks, must be duplicated because failure domains share nothing. However, since resources, such as disks, have become inexpensive, not sharing resources is less costly than previously and can greatly simplify creating fault tolerant service instances. Furthermore, because the primary and backup service instances share nothing, they can communicate with each other using only network protocols. It should be noted that hardware, on which services can be replicated, can be dissimilar, for example, but without limitation, “open” components and customized components, or some combination allowing greater applicability of the present invention.
The share-nothing nature of the service instances and communicating via network protocols is in contrast to conventional approaches where the primary and backup processes share resources, for example and without limitation disks, requiring that they maintain the same internal view of the shared resource they are controlling. Maintaining the same internal view requires checkpointing the internal state so that the primary and backup processes represent the internal states the same and can track the work each has done to the shared resource. Typically, checkpointing internal states in these conventional approaches requires internal communication across a bus or some other internal communication path that use of networking protocols. Since in an embodiment of the present invention only the same external view is maintained, the internal state is not tracked and checkpointing is vastly simpler.
Maintaining the same external view means that checkpointing is done for only external requests and messages that make changes to the database. Messages and requests, which return values that do not change regardless of the number of times the operation is performed, are not checkpointed. Any operations, which do not affect the state of the database, for example and without limitation, read operations, are not checkpointed. This is a departure from conventional methods, which checkpoint read operations because internal states must be represented in the same way.
In an embodiment of the present invention, internal states of the primary and its backups can have completely different representations. A primary can be implemented on top of one type of database while the backup can be implemented on top of another type of database. They can continue to communicate with each other using network protocols. The communication model employed by and the checkpointing method of an embodiment of the present invention (which does not perform checkpointing of read operations) will be explained in more detail below.
A client is a sequence of logically related requests, that is, a stream of communication. A client can be, for example and without limitation, a machine, a process on a machine, or a thread within a type of process.
A backup is slightly behind the primary in its processing because 1) the backup cannot start its work until after sending an acknowledgment to the primary and 2) the backup processes its work in the background. In addition, requests are independent and are committed in the primary and backup(s) separately. The difference in completion times is usually the time it takes the message to be sent from a primary to its backup and the backup to send the acknowledgment to the primary before starting its own processing. After the backup processes a request, the backup must commit its work to the database. Checkpointing is a form of logical mirroring, so that the logical state of the backup is a mirror copy of the primary. After a primary (102a-102N) commits the request or message to a database, the primary returns a response to the client (101a1-101bN) that sent the request or message (step 9).
Referring now to
Unlike conventional checkpointing methods, which allow concurrent operations at a primary and backup for processing messages and requests in a different order, an embodiment of the present invention typically has the backup commit the work for these transactions in order. To maintain the order in which the work is committed, the primary assigns a service sequence number to each request it receives from the client. As shown in
As shown, a backup could start executing the requests in any order if it has parallel threads. The order of execution is not important, however, for correct operations, the backup must commit the Create A request from Client 1 before committing the Delete A request from Client 2 as indicated by the service sequence number order. If the backup does not commit these requests in this order, an unintended result, for example, A would still exist after committing both requests instead of A no longer existing in the database after committing both requests. Attaching a service sequence number to each request and committing them in sequence obviates unintended results, that is, results caused by processing requests out of order. It should be noted that the backup can commit unrelated requests out of order and the logical state is the same as if the requests have been committed in order. However, on a failover, the backup must complete all pending requests before completing new requests to maintain the correct logical state.
In an embodiment of the present invention, to ensure correct ordering of uncommitted processed requests, the backup checks the service sequence numbers of the processed requests in its table to confirm that requests with lesser service sequence numbers have been committed (step 6). The backup commits requests in service-sequence-number order (step 7). The backup cannot commit the work to the database out of service-sequence-number order. Commits to the database at the backup would be blocked if the work associated with a request with a lower service sequence number has not been committed first. A commit to the database causes the relevant message table at the backup or primary to be updated with the response to the request (step 8). It should be noted that the same response maintained in the primary message table after the commits the work to its database (see step 5 above), is eventually sent to the client from the primary later in the process (see step 8 in FIG. 3A).
The basic fault tolerance model used in this checkpointing scheme is optimized. If a failure occurs after the client sends a request but before the response reaches the client, an independent mechanism rapidly detects the failure. The mechanism informs the client that it must resend the request and to which service instance (primary or backup) it must resend the request. This means that if any failure occurs, the client must close the old connection, make a new connection, resend the message, and wait for a response. If the request fails again, the client must try again.
Referring now to
Looking now to
It should be noted that if a client sends multiple requests before waiting for responses, a failure is handled the same. For example, if a client has five requests outstanding, that is, it has not received responses for five requests, the client must resend the requests. In the case of a path failure, the client sends the requests to the primary and the primary checks its table for responses. If the table contains responses to these requests, the primary resends the responses to the client. In the case of a primary failure, the client resends the requests to the backup and the backup checks its table for responses. If the table contains the responses, the backup sends the responses to the client. If the table contains the messages, but does not contain responses, the backup waits until its work for the requests are committed. As soon as they are committed and it updates its table with the responses, it sends the responses to the client.
Therefore, whether a client has a single outstanding request or multiple outstanding requests, the process is the same. The client must re-establish the connection to the current primary, which in some cases may be the previous backup. The client must then resend the message(s). When the current primary receives the request(s), it retrieves the response(s) either from its table, or if the request(s) has not been committed, processes the request(s) again. The primary then returns the response(s) to the client. However, since each response carries the request sequence number, the primary can return responses in any order.
An embodiment of the present invention also supports transaction processing, which includes handling message table entries and handling failures during transaction processing.
Looking now to
Alternatively, the primary can individually process table entries marked as part of a transaction. The primary checkpoints each request that is part of the transaction to the backup and commits the transaction. However, instead of the commit being a transaction commit to the database, the client's commit request is only added to the internal state of the transaction.
Looking now at
Regardless of the transaction checkpointing model used, so long as the transaction model is the same in both implementations (that is, both the primary and backup use batch or both use individual table entry transaction processing), the same external semantics are preserved including when the internal implementations of the transaction checkpointing model at the primary and backup are completely different. In addition, if a failure occurs during transaction processing (batch and individual) before the checkpoint, the transaction is aborted and the primary must restart the transaction for processing.
Similar to when operating under normal conditions, an embodiment of the present invention must implement checkpointing when the primary or backup is lost. In this event, a new machine must be brought up and made current to avoid a single point of failure. If a primary is lost, the backup becomes the primary and client requests are directed to it. A new empty backup that has no internal state is created on a different machine. The current primary maintains a request history table of service sequence numbers and for remote replication purposes also maintains a log of logical operations out on disk. When the old primary is lost, the new primary notes the starting point in the log and the service sequence number corresponding to that point in the log. The primary copies its current database to the backup, and then in effect, replicates all logical operations and service sequence numbers to the backup after the starting point in the log. The service sequence numbers act as a filtering mechanism to ensure data integrity, in particular, data elements copied at one point in time may be affected by data elements copied later in time. If no filtering is done, the data in the database may not reflect the most current data with all of its transformations.
It should be noted that the primary can leverage any replication mechanism that a computing system may have in place and is not limited to any particular replication mechanism to copy the logical operations and service sequence numbers to the backup.
After noting the starting point in the log, the primary then checks forward through the log for all newer operations that occurred since replication began and applies the operations to the backup. When the backup and primary contain nearly the same information, the primary stops accepting requests from clients in an effort to allow the backup to catch up with the primary's current processing. As soon as the backup has caught up with the primary's current state and the backup sends an acknowledgment, the primary can begin accepting new requests from the client and checkpoint them over to the backup as in a normal checkpointing scheme. Effectively, the primary is processing only new requests and the backup is processing new and old requests in the background until it catches up with the primary. However, straight copying over to the backup such as this can result in a “fuzzy copy”. An embodiment of the present invention overcomes the “fuzzy copy” problem by filtering out entries that have already been applied to the data that has been copied to the database.
An embodiment of the present invention provides a method for filtering out entries that have already been applied. The primary uses service sequence numbers as generation label for the data. As entries are placed in the message table with the service sequence number the data in the database also maintains the service sequence numbers as generation numbers. If an attempt is made to apply a change to a service object at the backup where the service sequence number is lower than in the data object, the backup learns that the entry has already been applied and skips the entry.
There is an optimization in the event of a soft failure. For example, if the backup is running on a machine that has a memory parity error or a memory error that is not recoverable, the machine will crash and reboot. However, the metadata on disk remains internally consistent because it is in a database, but it is obsolete because the primary did not send the updates while the backup was rebooting. The backup is missing log entries. In an embodiment of the present invention, if the backup experiences a soft failure, the backup communicates to the primary the last service sequence number it has in its log, that is, recorded on the backup's log on disk. The primary will behave send log entries from the point of the service sequence number. It should be noted that, this optimization occurs only if the primary has not recycled it message table. If the primary has recycled its table and the service sequence number is no longer in the primary's table, the backup will start over as if it were an empty backup being brought up for the first time. Therefore, if the backup experiences a quick reboot, refreshing its data can be done very quickly, applying only those entries with a service sequence number greater than the one communicated by the backup to the primary. If on the other hand, the backup is down for a few weeks and then comes back up, the backup will not updated with current information, instead it will be placed in the spare pool.
Although the present invention has been particularly described with reference to the preferred embodiments thereof, it should be readily apparent to those of ordinary skill in the art that changes and modifications in the form and details may be made without departing from the spirit and scope of the invention. It is intended that the appended claims include such changes and modifications. It should be further apparent to those skilled in the art that the various embodiments are not necessarily exclusive, but that features of some embodiments may be combined with features of other embodiments while remaining with the spirit and scope of the invention.
| Number | Name | Date | Kind |
|---|---|---|---|
| 4228496 | Katzman et al. | Oct 1980 | A |
| 4590554 | Glazer et al. | May 1986 | A |
| 4823256 | Bishop et al. | Apr 1989 | A |
| 5155678 | Fukumoto et al. | Oct 1992 | A |
| 5157663 | Major et al. | Oct 1992 | A |
| 5271013 | Gleeson | Dec 1993 | A |
| 5363503 | Gleeson | Nov 1994 | A |
| 5432926 | Citron et al. | Jul 1995 | A |
| 5455932 | Major et al. | Oct 1995 | A |
| 5621885 | Del Vigna, Jr. | Apr 1997 | A |
| 5684991 | Malcolm | Nov 1997 | A |
| 5802265 | Bressoud et al. | Sep 1998 | A |
| 5845328 | Maya et al. | Dec 1998 | A |
| 5968185 | Bressoud et al. | Oct 1999 | A |
| 6119129 | Traversat et al. | Sep 2000 | A |
| 6161193 | Garg et al. | Dec 2000 | A |
| 6256751 | Meth et al. | Jul 2001 | B1 |
| 6275953 | Vahalia et al. | Aug 2001 | B1 |
| 6584477 | Mosher, Jr. | Jun 2003 | B1 |
| Number | Date | Country |
|---|---|---|
| 40-5055934 | Mar 1993 | JP |
| Number | Date | Country | |
|---|---|---|---|
| 20040199812 A1 | Oct 2004 | US |
