Not Applicable
A portion of the material in this patent document is subject to copyright protection under the copyright laws of the United States and of other countries. The owner of the copyright rights has no objection to the facsimile reproduction by anyone of the patent document or the patent disclosure, as it appears in the United States Patent and Trademark Office publicly available file or records, but otherwise reserves all copyright rights whatsoever. The copyright owner does not hereby waive any of its rights to have this patent document maintained in secrecy, including without limitation its rights pursuant to 37 C.F.R. § 1.14.
Not Applicable
1. Field of the Invention
The invention relates to software-based fault-tolerant computer systems and, in particular, to multithreaded application programs that are replicated using the leader-follower semi-active and passive replication strategies.
2. Description of Related Art
Fault-tolerant systems are based on entity redundancy (replication) to mask faults and, thus, to provide continuous service to their users. In software fault tolerance, the entities that are replicated are the application programs or parts thereof (processes, objects or components). A fundamental issue in the design and implementation of fault-tolerant systems is that of maintaining the consistency of the states of the replicas.
Distributed systems offer the opportunity for fault tolerance by allowing replicas of the application programs to be hosted on different computers (i.e., in different fault containment regions). In the client-server model of distributed computing, a client invokes a method of a server, typically hosted on a different computer, by receiving a request message containing that method invocation and by receiving a reply message from that server. To render an application fault-tolerant, the server is replicated but the client may also be replicated, particularly in multi-tier and peer-to-peer applications, where a process, object or component acts as both a client and a server.
Fault-tolerant systems support several different replication strategies including semi-active and passive replication, and variations thereof. In both semi-active and passive replication, one of the replicas is distinguished as the Primary replica and the other replicas are called the Backup replicas.
In semi-active replication, all of the replicas of a process, object or component execute each method invoked on the replicas. The Primary replica determines the order in which the methods and other operations are executed and communicates that order to the Backup replicas, which executes the methods and other operations in the same order. If the Primary replica makes a decision regarding a non-deterministic operation (such as the order in which access to a shared resource is granted), it communicates that decision to the Backup replicas which make the same decision. If the Primary replica fails, a Backup replica takes over as the Primary replica and starts making decisions that the other Backup replicas must follow.
In passive replication, only the Primary replica executes the methods invoked on the replicas. The state of the Primary replica (values of its variables or attributes) is checkpointed periodically or on demand, and the messages, methods and other operations after the checkpoint are logged. If the Primary replica fails, a Backup replica takes over as the Primary replica. The checkpoint is loaded into the Backup replica and the messages, methods and other operations after the checkpoint are replayed.
A challenging aspect of replication is to maintain strong replica consistency, as methods are invoked on the replicas and the states of the replicas change dynamically, and as faults occur. Strong replica consistency means that, for each method invocation or operation, for each data access within that method invocation or operation, the replicas obtain the same data values. Moreover, for each result, message sent or request made to other processes, objects or components, the replicas generate the same result, message or request.
Many application programs written in modern programming languages (such as C++, Java, etc.) involve multithreading, which is a source of non-determinism. Unless it is properly handled, non-determinism can lead to inconsistency in the states of the replicas. To maintain strong replica consistency, it is necessary to sanitize or mask such sources of non-determinism, i.e., to render a replicated application program virtually deterministic. A virtually deterministic replicated application program is an application program that exists as two or more replicas and that may involve non-deterministic decisions; however, for those non-deterministic decisions that affect the states of the replicas, the replicas must make the same non-deterministic decisions.
U.S. Pat. Nos. 5,577,261 and 5,794,034 which are incorporated herein by reference describe the implementation of “process management” functions, such as the claim( ), release( ), suspend( ) and signal( ) functions, which are also used by the current invention. Operations involving those methods are rendered consistent by having each processor claim a global mutex (called GLUPP) before performing any “process management” operation. Once it has acquired the global mutex, the process performs the operation and then distributes the results to the other processors before relinquishing the global mutex.
U.S. Pat. No. 4,718,002 which is incorporated herein by reference describes how a mutex can be granted to processors, processes, replicas or threads in a distributed system. Each grant of a mutex requires three messages, two messages to claim and grant the mutex and one message to release the mutex. It should be appreciated that this approach requires the communication of multiple additional messages for claiming, granting and releasing a mutex.
U.S. Pat. No. 5,621,885 which is incorporated herein by reference describes a strategy based on Primary/Backup replication, in which the Primary replica executes the required operations. When the Primary replica performs an I/O operation, the results of the I/O operation are communicated to the Backup replica, so that the Backup replica performs the same operation as the Primary replica. This strategy is directed at maintaining consistency between Primary and Backup replicas only for I/O operations and does not address inconsistency that arises from multithreading.
U.S. Pat. Nos. 5,802,265 and 5,968,185 which are incorporated herein by reference are related to the TFT system described below and describe a strategy based on the Primary/Backup approach, in which the Primary replica executes the required operations. When the Primary replica performs an asynchronous or non-deterministic operation, it communicates the results of that operation to the Backup replica, so that the Backup performs the same operation as the Primary. The teachings of these patents disclose no mechanism for guaranteeing that a Backup receives such communication before or concurrently with the communication of results by the Primary to an entity external to the system. As a result, the design is exposed to the risk that the Primary might perform actions and communicate results of those actions to clients, and subsequently fail without ensuring that the Backups have received the communication from the Primary about the operating system interactions. It should be appreciated that such a fault can leave a Backup with the obligation of reproducing those actions; however, the Backup replica might lack the necessary information to do so.
The TARGON/32 system (A. Borg, W. Blau, W. Graetsch, F. Herrmann and W. And, Fault tolerance under Unix, ACM Transactions on Computer Systems, vol. 7, no. 1, 1989, pp. 1-24, incorporated herein by reference) provides mechanisms for the Unix operating system that ensure consistent processing by multiple replicas of asynchronous operations and signals, such as the claim( ) and release( ) functions. A designated control processor (the Primary) records a checkpoint immediately before it processes an asynchronous operation. If the control processor fails, a Backup processor restarts from the checkpoint and then processes the asynchronous operation immediately thereafter, ensuring that the Backup processes the operation starting from the same state as the control processor.
The Delta-4 system (D. Powell (ed.), Delta-4: A Generic Architecture for Dependable Distributed Computing, Springer-Verlag, 1991, incorporated herein by reference) provides support for non-deterministic application programs that employ semi-active or passive replication. To provide such support, Delta-4 queues interrupts until the application program executes a polling routine in which the replicas synchronize and agree on the interrupts received and the order in which to process them.
The Hypervisor system (T. C. Bressoud and F. B. Schneider, Hypervisor-based fault tolerance, ACM Transactions on Computer Systems, vol. 14, no. 1, 1996, pp. 80-107, incorporated herein by reference) and the Transparent Fault Tolerance (TFT) system (T. C. Bressoud, TFT: A software system for application-transparent fault tolerance, Proceedings of the IEEE 28th Fault-Tolerant Computing Symposium, Munich, Germany, June 1998, pp. 128-137, incorporated herein by reference) uses a Primary/Backup approach and aims for transparency to the application and the operating system by utilizing hardware instruction counters to count the instructions executed between two hardware interrupts. The TFT system utilizes object code editing to modify the program code to provide fault tolerance.
Other researchers (J. H. Sly and E. N. Elnozahy, Supporting non-deterministic execution in fault-tolerant systems, Proceedings of the IEEE 26th Fault Tolerant Computing Symposium, Sendai, Japan, June 1996, pp. 250-259, incorporated herein by reference) have introduced a software instruction counter approach, analogous to the hardware instruction counter approach of the Hypervisor system, to count the number of instructions between non-deterministic events in log-based rollback-recovery systems. If a fault occurs, the instruction counts are used to replay the instructions and the non-deterministic events at the same execution points.
Non-preemptive deterministic scheduler strategies also exist that impose a single logical thread of control on the replicas to maintain strong replica consistency (P. Narasimhan, L. E. Moser and P. M. Melliar-Smith, Enforcing determinism for the consistent replication of multithreaded CORBA applications, Proceedings of the IEEE 18th Symposium on Reliable Distributed Systems, Lausanne, Switzerland, October 1999, pp. 263-273, incorporated herein by reference). The effect of this strategy is to undo the multithreading that was programmed into the application program.
Transactional Drago (S. Arevalo, R. Jimenez-Peris and M. Patino-Martinez, Deterministic scheduling for transactional multithreaded replicas, Proceedings of the IEEE 19th Symposium on Reliable Distributed Systems, Nurnberg, Germany, October 2000, pp. 164-173, incorporated herein by reference) also uses a non-preemptive deterministic scheduler but is configured for use in transaction processing systems.
Therefore, a need exists for systems, software mechanisms, methods, improvements and apparatus for providing strong replica consistency for multithreaded application programs based on semi-active and passive replication that maintain application transparency. The systems, software mechanisms, methods, improvements and apparatus in accordance with the present invention satisfy that need, as well as others, and overcome deficiencies in previously known techniques.
The present invention can be described as a system, software mechanisms, methods, improvements and apparatus that aim at achieving strong replica consistency of multithreaded application programs that are replicated using the leader-follower (primary-backup) semi-active or passive replication strategy. This invention is applicable to distributed systems in which the several computers within the distributed system share no memory and communicate with each other by messages.
Unlike U.S. Pat. Nos. 5,802,265 and 5,968,185, described above, the present invention ensures that the Backup replicas receive the necessary access ordering information before or concurrently with the communication of results by the Primary replica to an entity external to the system. Unlike the Hypervisor system, previously described, the current invention does not attempt to maintain transparent fault tolerance by counting instructions executed between hardware interrupts. Moreover, unlike U.S. Pat. No. 4,718,002, described above, the present invention does not require the communication of additional separate messages for claiming, granting and releasing a mutex over and above the regular messages that are transmitted during normal operation. Unlike the non-preemptive scheduler strategies, previously described, the current invention allows for the maximum degree of concurrency of threads while maintaining strong replica consistency.
An aspect of this invention is to provide mechanisms for fault-tolerant systems based on replication, in which an application process, object or component is replicated, and the replicas of the application process, object or component are typically located on different computers within a distributed system. In the event that one of the replicas is disabled by a fault, another replica can continue to provide service.
Another aspect of this invention is to employ semi-active or passive replication, in which there are two or more replicas of an application process, object or component, one of which is distinguished as the Primary replica and the others of which are called the Backup replicas. If the Primary replica is disabled by a fault, one of the Backup replicas takes over as the Primary replica and continues to provide the service.
There are many sources of nondeterminism in application programs. The mechanisms of this invention address nondeterminism caused by multithreading in replicated application programs that use the semi-active or passive replication strategy. They assume that the application program has been correctly coded so that each resource that is shared by two or more threads in a process, object or component is protected by a shared resource control construct, such as a mutex, that allows only of those threads to access the shared resource at a time. They assume further that each thread has a unique thread identifier and that each mutual exclusion construct (mutex) has a unique mutex identifier.
To maintain strong replica consistency for replicated multithreaded application programs, it is necessary to sanitize or mask sources of non-determinism. The mechanisms of the present invention sanitize multithreaded application programs based on semi-active or passive replication strategies. They mask multithreading as a source of non-determinism in the replicas, so that strong replica consistency is maintained.
The present invention exploits a reliable source-ordered multicast protocol to maintain strong replica consistency. Reliable means that the replicas of a process, object or component receive the messages that are multicast to it. Source ordered means that the replicas of a process, object or component receive the messages from a given source in the same order.
For unithreaded application programs based on semi-active or passive replication, a multicast protocol that delivers request and reply messages reliably and in the same order to the replicas can be used to maintain strong replica consistency. For multithreaded application programs, the problem of maintaining strong replica consistency is more difficult because two threads in a replica can access shared resources, such as shared data, in an order that is different from the order in which the corresponding threads in another replica access the shared data; consequently, the states of the replicas can become inconsistent.
For multithreaded application programs, if two threads access data that are shared between them, then only one of those threads can access that shared data at a time. Therefore, to maintain strong replica consistency, the shared data must be protected, preferably with a mutex, wherein a thread must claim the mutex and enter a critical section of code, preferably protected with a mutex, before accessing the shared data. In the presence of replication, the threads in the replicas must claim the mutexes, enter the critical sections and access the shared data in an identical order.
The mechanisms of this invention ensure that, for replicated multithreaded application programs based on the leader-follower (passive-backup) strategy of semi-active or passive replication, the threads in the replicas are granted their claims to mutexes, semaphores and so forth in the same order, even though the threads in the replicas might claim the mutexes, semaphores and so forth in different orders. As a result, the mechanisms of this invention eliminate multithreading as a source of nondeterminism in replicated multithreaded application programs. In other words, they sanitize such programs by rendering those programs virtually deterministic.
The mechanisms of this invention depend on the existence of a Primary replica that dictates the order in which the threads in the Backup replicas claim, and are granted, accesses to shared resources (i.e., mutexes). The Primary replica piggybacks this resource access ordering and granting information (i.e., mutex ordering information) on the messages that it multicasts.
In the Primary replica, the mechanisms of the present invention piggyback access ordering information, such as mutex ordering information, onto the regular messages that they multicast. The access ordering information conveys the order in which the threads in the Primary replica have been granted their claims to access shared resources. In the Backup replicas, the mechanisms of the present invention ensure that the corresponding threads are granted corresponding claims to access the shared resources, in the same order as the threads in the Primary replicas, as dictated by the access ordering information received from the Primary replica.
Embodiments of the present invention are exemplified with threads of replicas claiming and releasing mutexes, semaphores and so forth, associated with shared resources, preferably shared data and code that manipulates such shared data. In a Backup replica, the granting of a given resource, to a given thread, for a given claim number, is constrained until the Primary has been granted the corresponding access, and other accesses corresponding to previous accesses have been completed by the Backup replica, and has communicated those accesses to the Backup replica. In this way, threads in Backup replicas are constrained to follow the order and granting of shared resource accesses by threads in the Primary replica, as communicated by the Primary replica to the Backup replicas. As a result, the system, methods, software mechanisms, improvements and apparatus of the present invention sanitize replicated multithreaded application programs, by masking multithreading as a source of non-determinism and, thus, render such programs virtually deterministic, thereby maintaining strong replica consistency.
To maintain application transparency for replicated multithreaded application programs based on the leader-follower strategy of semi-active or passive replication, while maintaining strong replica consistency, the invention is preferably implemented using the technique of library interpositioning to intercept the calls to functions in the operating system's thread library and to divert them to another library. An example of this other library is herein referred to as the Consistent Multithreading (CMT) library, which is the preferred embodiment of this invention and is so-named because it eliminates multithreading as a source of non-determinism in replicated multithreaded application programs and maintains strong replica consistency.
The Consistent Multithreading (CMT) library is interposed ahead of the operating system's thread library, such as the standard POSIX thread (PTHREAD) library. The CMT library contains control program code comprising wrapper functions for the functions of the operating system's thread library that claim and release mutexes, semaphores, condition variables, and so forth. The application program invokes the wrapper functions of the CMT library, instead of the corresponding functions of the operating system's thread library. The wrapper functions of the CMT library subsequently invoke the corresponding functions of the operating system's thread library. This allows the CMT library to modify the behavior of the replicated multithreaded application program, without modifying either the application program or the functions of the operating system's thread library.
When a thread in the Primary replica invokes a function to claim a mutex, the CMT claim( ) function is invoked. The CMT claim( ) function in turn invokes the claim( ) function of the operating system's thread library and subsequently piggybacks ordering and granting information onto the next message that it multicasts. The ordering and granting information specifies which thread in the Primary replica has been granted access to a shared resource (mutex) for a particular access request. The multicast protocol delivers messages reliably and in the same source order to the Backup replicas. In the Backup replicas, the threads are granted accesses to shared resources (mutexes) in the same order as the corresponding accesses were granted to the corresponding threads in the Primary replica, based on the ordering and granting information that Primary replica multicasts to the Backup replicas.
The multicast protocol might communicate a message containing the ordering and granting information directly to the Backup replicas. Alternatively, and preferably, the ordering and granting information is piggybacked on the regular messages. If a regular message, such as a reply message sent by a Primary server replica to a client in response to a request message from that client, can be multicast to both the client and the Backup server replicas, then the regular message communicates the ordering and granting information to the Backup server replicas. If, however, the multicast protocol is such that the regular message from the Primary server replica cannot be multicast to both the client and the Backup server replicas, then the ordering and granting information is piggybacked on the next regular message that the Primary server sends to the client, and the client also piggybacks that same ordering and granting information on the next message it multicasts, either its next request message or an acknowledgment message for the reply message, to the Primary and Backup server replicas.
The mechanisms of this invention sanitize the non-deterministic behavior of the threads in the different replicas in claiming mutexes, in a manner that is transparent to the application program, and similarly for the release( ) function and for semaphores, control variables, and other similar structures.
For application programs that run on an operating system that provides Dynamically Linked Libraries (DLL) (e.g., Solaris, Linux, Windows), a command is issued to the DLL mechanisms that causes the DLL mechanisms to interpose the Consistent Multithreading (CMT) library, containing the wrapper functions, ahead of the operating system's thread library. This interpositioning causes the application program to invoke the functions of the CMT library, rather than the corresponding functions of the operating system's thread library directly. Thus, the mechanisms involve no modification or recompilation of the application program and, thus, are transparent to the application program.
If, on the other hand, the operating system does not provide Dynamically Linked Libraries (e.g., VxWorks), it is necessary to insert a statement into the makefile for the application program that directs the linker to include the CMT library ahead of the operating system's thread library. In this case, the application program is not modified but the makefile is modified.
The mechanisms of this invention allow concurrency of threads that do not simultaneously claim the same mutex, or that claim different mutexes, while maintaining strong replica consistency. As a result, the mechanisms of the invention provide the maximum degree of concurrency while maintaining strong replica consistency.
The mechanisms of this invention sanitize multithreaded application programs in that they mask multithreading as a source of non-determinism so that strong replica consistency is maintained.
Thus, the present invention provides an extension to existing systems, in particular that have an operating system thread library that grants access to shared resources without regard to consistent granting of those accesses across multiple replicas, to provide fault tolerance and consistent replication based on the semi-active or passive replication strategies. It should be appreciated that, in the current invention, the Primary replica dictates the order in which threads in a Backup replica are granted accesses to shared resources, which order is the same as the order in which the corresponding threads in the Primary replica are granted corresponding accesses to corresponding shared resources. Further aspects of the invention include: (1) at the Primary replica, piggybacking access ordering information onto regular multicast messages that specifies the order in which threads in the Primary replica have been granted accesses to shared resources, and (2) at a Backup replica, receiving multicast messages from the Primary replica that contain the access ordering information, which determines the order in which corresponding threads in the Backup replica are granted corresponding accesses to corresponding shared resources.
The source-ordered multicast protocol preferably piggybacks the ordering information on regular messages being multicast, therein minimizing communication overhead. The invention ensures that the Backup replicas have the access ordering information that they need, in the event that the Primary replica fails, to order to maintain strong replica consistency.
Further aspects of the invention will be brought out in the following portions of this document, wherein the detailed description is for the purpose of fully disclosing preferred embodiments of the invention without placing limitations thereon.
The invention will be more fully understood by reference to the following drawings which are for illustrative purposes only:
Referring more specifically to the drawings, for illustrative purposes the present invention is embodied in the apparatus generally shown in
It should be noted that the mechanisms of the invention pertain to the semi-active and passive replication strategy. The algorithms, diagrams and descriptions below generally refer to semi-active replication, but the mechanisms apply equally to passive replication. Only one Backup replica is shown in the diagrams, but the mechanisms of the invention apply equally if multiple Backup replicas are used.
In the algorithms, diagrams and descriptions below, the term mutex refers to a mutual exclusion construct, in particular a binary semaphore, which respects the priority transfer mechanisms of the scheduler for an operating system. However, the mechanisms of the invention apply equally well to counting semaphores, condition variables, and other similar mechanisms for controlling access to a shared resource.
The mechanisms of this invention employ the technique of library interpositioning to intercept the calls to functions in the operating system's thread library. For example, for the standard POSIX Thread (PTHREAD) library, the mechanisms intercept the calls to the functions in the PTHREAD library. The Consistent Multithreading (CMT) library that is the preferred embodiment of this invention contains wrappers for the functions in the operating system's thread library, such as the pthread_mutex_lock( ) and pthread_mutex_unlock( ) functions in the PTHREAD library. In view of the applicability of these mechanisms to thread libraries, other than the PTHREAD library, these functions and the corresponding wrapper functions are referred to more generally as claim( ) and release( ), respectively.
When a thread of the Primary replica invokes a function to claim a mutex, the claim( ) wrapper function of the CMT library is invoked, instead of the corresponding claim( ) function in the operating system's thread library. The CMT claim( ) function takes appropriate steps, shown in
For each mutex M, or other form of shared resource control, the consistent multithreading mechanisms maintain a boolean variable, such as M.available, which indicates whether or not mutex M is available. For each thread T, the consistent multithreading mechanisms maintain a boolean variable, such as T.suspended, which indicates whether or not thread T is suspended.
In the Primary replica, when thread T claims a mutex M as its Nth claim to any mutex, and that mutex is already being held by another thread, the consistent multithreading mechanisms set T.suspended to true and suspend thread T using the standard mutex scheduling mechanisms of the operating system, which subsequently resume thread T, using priority transfer, as appropriate.
In the Primary replica, the consistent multithreading mechanisms maintain a queue, the ordered claims queue of triples (T, M, N), where thread T has claimed, and has been granted, mutex M and this is the Nth claim for any mutex that thread T has made. The ordered claims queue spans different threads, different mutexes and different claims by the threads. It should be appreciated that the Nth claim numbering can be based on the numbering of claims to any mutex, numbered claims for mutexes from a given thread, numbered claims for claims to a given mutex from a given thread, or otherwise identified according to other alternative means for identifying claims.
In the Primary replica, as the mutexes are granted to the threads, the consistent multithreading mechanisms append a triple (T, M, N) to the ordered claims queue, where N is the Nth claim of any mutex by thread T. The consistent multithreading mechanisms remove one or more such triples from the ordered claims queue, and piggyback those triples (the mutex ordering information) onto the next message that they send, as shown at 100 and 114 in
In a Backup replica, thread T is not allowed to claim mutex M as its Nth claim to any mutex, until the message from the Primary replica that contains the ordering information (T, M, N) is ordered and delivered to the Backup replica. The underlying reliable source-ordered multicast protocol guarantees that the Backup replicas receive the messages from the Primary replica and that they receive those messages in the same order.
In a Backup replica, for each mutex M, the consistent multithreading mechanisms maintain an M.orderedClaims queue of (T, N) pairs, in the order in which the Backup replicas receive those claims from the Primary replica. The consistent multithreading mechanisms at a Backup replica append a (T, N) pair to the M.orderedClaims queue, as the Backup replica receives messages containing the (T, M, N) triples from the Primary replica.
In a Backup replica, when thread T claims a mutex M as its Nth claim to any mutex, if the entry (T, N) is not the first entry in the M.orderedClaims queue, the consistent multithreading mechanisms set T.suspended to true. The consistent multithreading mechanisms resume a thread T that was suspended in the order in which the pairs (T, N) occur in the M.orderedClaims queue, rather than in the order in which they were suspended or an order determined by the operating system scheduler.
Two Multithreaded Replicas with the Interposed CMT Library
By way of example,
In the Primary replica 12 there are four threads 16, comprising threads T1-T420-26. Thread T120 and thread T222 access the shared resource 38 comprising the shared data D144 using mutex M146, which protects that shared data. Thread T324 and thread T426 access the shared resource 40 comprising the shared data D248 using mutex M250, which protects that shared data. The CMT library 42 is interposed between the Primary replica 16 and the operating system's thread library and is dynamically linked into the Primary replica.
Similarly, in the Backup replica 14 there are four corresponding threads 18, comprising threads T1-T428-34. Thread T128 and thread T230 access the shared resource 52 comprising the shared data D158 using mutex M160, which protects that shared data. Thread T332 and thread T434 access the shared resource 54 comprising the shared data D262 using mutex M264, which protects that shared data. The CMT library 56 is interposed between the Backup replica 18 and the operating system's thread library and is dynamically linked into the Backup replica.
In the Primary replica 12, because thread T120 and thread T222 can each read and write the shared data D144, access to which is protected by mutex M146, and, similarly, for thread T324 and thread T426 and the shared data D248. However, thread T120 and thread T222 share no data with thread T324 and thread T426; as a result thread T120 and thread T222 can execute concurrently with thread T326 and thread T428 without the need for a mutex. The same is true for the corresponding threads in the Backup replica.
Scenario That Illustrates the CMT Mechanisms
Similarly, a Backup replica 72 executes threads T180, T282 and T384. Thread T180 and thread T282 access shared data, protected by mutex M 88. Thread T384 does not share data with thread T180 and thread T282 and, thus, executes concurrently with them without the need for a mutex.
In the Primary replica 70, when thread T174 invokes the CMT claim( ) function to claim mutex M, the CMT claim( ) function invokes the claim( ) function of the operating system's thread library to claim mutex M 86. If mutex M 86 is unavailable, the operating system suspends thread T174. The operating system may use thread priority to determine the thread to which it grants the mutex next, and may use priority inheritance to expedite the granting of mutexes to high priority threads. Referring to the figure, because mutex M 86 is available and thread T174 is the highest priority thread waiting for mutex M, the operating system grants mutex M 86 to thread T174. The CMT claim( ) function then appends the claim (T1, M, 8) 94 to the M.orderedClaims queue 92, and returns to the application program, granting mutex M 86 to thread T174 and allowing it to proceed. Subsequently, the consistent multithreading mechanisms extract the claim (T1, M, 8) from the M.orderedClaims queue, which then becomes empty 102, and multicast a message containing the claim (T1, M, 8) piggybacked onto the message 100.
Similarly, when thread T276 in the Primary replica 70 invokes 104 the claim( ) function of the CMT library to claim mutex M 86, the CMT claim( ) function invokes the claim( ) function of the operating system's thread library to claim mutex M 86. Because mutex M 86 is unavailable, having been granted to thread T174, the operating system suspends 106 thread T276.
When thread T174 invokes 108 the release( ) function of the CMT library to release mutex M 86, the CMT release( ) function invokes the release( ) function of the operating system's thread library to release mutex M 86, and then returns, allowing thread T174 to proceed. The operating system grants 112 mutex M 86 to thread T276 and awakens that thread, which was previously suspended 106 by the CMT claim( ) function, as described above. It resumes the processing of that function, which then appends the claim (T2, M, 5) to the M.orderedClaims queue 110, and returns to the application program, granting 112 mutex M 86 to thread T276, and allowing it to proceed. Subsequently, the consistent multithreading mechanisms extract the claim (T2, M, 5) from the M.orderedClaims queue, which then becomes empty 116, and multicasts a message containing that claim piggybacked onto the message 114.
In the Backup replica 72, when thread T282 invokes 118 the claim( ) function to claim mutex M 88, the CMT claim( ) function checks whether the claim is the first entry in the M.orderedClaims queue 120. Because that queue is empty 122, the CMT claim( ) function sets T2.suspended 124 to true and then invokes the suspend( ) function of the operating system's thread library to suspend thread T282.
When the Backup replica 72 receives the claim (T1, M, 8) that the Primary replica 70 multicast 100, and that the multicast protocol delivered to the Backup replica 72, the consistent multithreading mechanisms in the Backup replica 72 appends (T1, 8) to the M.orderedClaims queue 126. Even though (T1, 8) is the first entry in the M.orderedClaims queue, because T1.suspended 358 in
When thread T180 in the Backup replica 72, subsequently invokes 128 the claim( ) function to claim mutex M 88, the CMT claim( ) function checks whether the claim (T1, 8) is the first entry in the M.orderedClaims queue and whether mutex M 88 is available. Because claim (T1, 8) is the first entry in that queue 130 and mutex M 88 is available, the CMT claim( ) function sets M.available to false 132 and then invokes the claim( ) function of the operating system's thread library to claim mutex M 88 for thread T180. The CMT claim( ) function then removes (T1, 8) from the M.orderedClaims queue, which then becomes empty 134, and returns, granting 136 mutex M 88 to thread T180.
The Backup replica 72 receives the claim (T2, M, 5) 138 that the Primary replica 70 multicast 114 and the multicast protocol delivered. The CMT message handler appends (T2, 5) to the M.orderedClaims queue 138. Even though (T2, 5) is the first entry in the M.orderedClaims queue, because M.available 356 in
When thread T180 in the Backup replica 72 has finished accessing the shared resource, it invokes 140 the release( ) function to release mutex M 88. The CMT release( ) function invokes the release( ) function of the operating system's thread library to release mutex M 88, sets M.available to true, and checks the M.orderedClaims queue. Because (T2, 5) is the first entry 142 in the M.orderedClaims queue (which indicates that thread T282 is waiting for mutex M 88), and because thread T2.suspended is true, the CMT release( ) function invokes the signal( ) function of the operating system's thread library to awaken thread T282. The release( ) function then returns, allowing T180 to proceed.
Meanwhile, thread T282 in the Backup replica 72, which was suspended by the CMT claim( ) function, awakens and resumes processing. The CMT claim( ) function checks M.available and also the M.orderedClaims queue. As (T2, 5) is the first entry in the M.orderedClaims queue, the CMT claim( ) function sets T2.suspended to false and sets M.available to false. The CMT claim( ) function then invokes the claim( ) function of the operating system's thread library to claim mutex M 88 for thread T282, and the operating system grants 144 mutex M 88 to thread T282. The CMT claim( ) function removes (T2, 5) from the M.orderedClaims queue, which then becomes empty 146. The CMT claim( ) function then returns, allowing thread T282 to proceed.
Mutex Ordering Requirements for Strong Replica Consistency
One might conclude that, for semi-active and passive replication, if the Primary server replica fails and will never subsequently generate a reply, then a Backup server replica can simply start processing the request that the Primary replica failed to complete, and now only the Backup replica matters. That conclusion is false. During its operation, before it failed, the Primary replica might have invoked methods of other servers. If a Backup replica processes a method invocation that the failed Primary replica did not complete, but the Backup replica invokes the methods of the other servers in a different order than the Primary replica invoked the methods before it failed, then inconsistencies in the states of the replicas can arise. Consequently, the Primary replica must communicate ordering information to the Backup replicas so that, in the event that the Primary replica fails, the threads of the Backup replicas can process the same requests in the same order, and can be granted the same mutexes in the same order, so that the processing by the Backup replica exactly duplicates the processing that the Primary replica performed before it failed, insofar as that processing is visible to other objects.
In the figure it can be seen that thread A processes a second request 168. Again, mutexes are granted in the order, thread B 188 followed by thread A 190. Thread A 154 then invokes a method of another server and generates a request 192. To maintain the consistency of the Backup replica 152, so that it can make that same request if the Primary replica 150 fails, the Backup replica 152 is provided the mutex ordering information prior to that request, and that information is preferably piggybacked 194 on the request message 192. After thread A 154 has received the reply 196 to its request 194, the mutex is granted first to thread B 198 and then to thread A 200 before thread A 154 generates its reply 202. It should be appreciated that mutex ordering information is piggybacked 204 on the reply message 202.
Thread A then starts to process another request 170. The mutex is granted to thread A 206 and then to thread B 208. Before thread A can complete its processing of this request, the Primary replica 150 fails 210 and 212. It did not communicate the mutex ordering information to the Backup replica; however, it also did not communicate any results of that processing to any other object. Thus, the order in which the mutex was granted is not significant.
The right side of
Similarly, in the Backup replica 152, during the processing of the next request 174, leading to the generation of a request 232, a reply 236 to that request and the reply to the client 238, the mutex is granted in exactly the same order 224, 226, 228, and 230 as that mutex was granted in the Primary replica 150. However, when thread A 160 processes the next request 176, no mutex ordering information is available. Consequently, in the Backup replica the mutex is granted to the threads in whatever order the operating system deems appropriate. That order 242, 244 differs from the order in which the mutex was granted to the threads in the Primary replica, and the reply generated by the Backup replica 246 might be different from the reply that the Primary replica would have generated had it not failed but, because the Primary replica became faulty and did not generate that reply, any such difference is of no significance.
Mechanisms That Communicate Mutex Ordering Information
The mechanism of
In general, it may be considered that message delivery in this case comprises delivering of two multicast messages, one message multicast by a first Primary replica to the replicas of other processes, objects or components and one message multicast by a second Primary replica of the other processes, objects or components to the first Primary replica and its Backup replicas
The mechanism depicted in
If the Primary server 254 fails after transmitting its reply message to the client, and the Backup servers (a single Backup server 256 shown by example while typically multiple Backup servers would exist) do not receive the reply, the Backup servers request the client replicas to send the most recent messages that they received from the Primary server 254 before it failed; thus, the Backup servers 256 obtain the mutex ordering information that they need to ensure that their processing is consistent with that reply.
A similar sequence of messages communicates the mutex ordering information to the Backup servers if the Primary server acts as a client and invokes a method of another server.
As shown in
Primary Replica Thread Invokes CMT Claim( ) Function to Claim a Mutex
In the Primary replica, when thread T invokes the CMT claim( ) function to claim mutex M, the CMT claim( ) function executes the following steps:
When the claim( ) function of the operating system's thread library grants mutex M to thread T, which could be sometime later, the CMT claim( ) function appends the triple (T, M, N) to the ordered claims queue of claims to be multicast as depicted in block 306 and then returns to the application program, granting mutex M to thread T and allowing T to proceed as per block 308. If mutex M is held by another thread, thread T will remain suspended until that other thread releases the mutex and the operating system's thread library grants mutex M to thread T.
Primary Replica Multicasts a Message with Piggybacked Ordered Claims
Periodically, or when they transmit a regular message, or when they add an entry to the ordered claims queue, the mechanisms in the Primary replica piggyback the entries of the ordered claims queue onto the message and multicast it, as shown at 98, 100, 114 in
Primary Replica Thread Invokes CMT Release( ) Function to Release a Mutex
When thread T of the Primary replica invokes the CMT release( ) function to release mutex M, the CMT release( ) function executes the following steps:
Backup Replica Invokes CMT Claim( ) Function to Claim a Mutex
At a Backup replica, when thread T invokes the CMT claim( ) function to claim a mutex M as its Nth claim of any mutex, the CMT claim( ) function executes the following steps:
It then checks whether (T, N) is the first entry in the M.orderedClaims queue 320. If (T, N) is the first entry, it sets M.available to false as per block 322 and invokes the claim( ) function of the operating system's thread library to claim mutex M for thread T as depicted at block 324.
When thread T is granted its claim of mutex M, the CMT claim( ) function removes (T, N) from the M.orderedClaims queue as block 326 and returns at block 328. Otherwise, if (T, N) is not the first entry in the M.orderedClaims queue, the CMT claim( ) function sets T.suspended to true as represented by block 330 and invokes the suspend( ) function of the operating system's thread library to suspend thread T as per block 332.
Backup Replica Invokes the CMT Release( ) Function to Release a Mutex
When thread T invokes the CMT release( ) function to release mutex M, the CMT release( ) function executes the following steps:
Backup Replica Receives Ordered Claims Piggybacked on a Message
When a Backup replica receives ordered claims piggybacked on a message, the consistent multithreading message handler extracts, in order, each ordered claim (T, M, N) from the message and then executes the following steps:
Backup Replica Thread is Awakened
When a thread T is awakened while waiting for mutex M as its Nth claim of any mutex, the CMT claim( ) function executes the following steps:
As can be seen, therefore, The present invention provides a system, software mechanisms, methods improvements and apparatus for determining the order in which shared resources are accessed by threads in multiple replicas, in a computer environment in which multithreaded application programs are replicated using the semi-active or passive replication strategies which are based on the leader-follower (primary-backup) approach. The invention generally exemplifies generating an order in which claims to mutual exclusion constructs (mutexes) are granted to threads in the Primary replica, and is communicated to the Backup replicas for constraining the granting of resource claims, so as to accord with the grant order of the Primary replica. Numerous aspects of the invention, preferred embodiments, and optional aspects are described. It will be appreciated that the present invention may be practiced in various embodiments, such as having differing levels of complexity and utilizing different routines and structures while still adhering to the patentable aspects of the present invention.
As can also be seen, the system, software mechanisms, methods improvements and apparatus for maintaining strong replica consistency are readily applied to both new and existing applications. They are applicable to computing environments that can execute replicated multithreaded application programs and are particularly well-suited to distributed systems having multiple computers, hosting Primary and Backup replicas, connected by a local-area or wide-area network. The system, software mechanisms, methods, improvements and apparatus may be used to replicate processes, objects, components and other software entities. In reference to software, the present system, software mechanisms, methods, improvements and apparatus may be supplied as a computer readable media, or other distribution form, containing program code and/or data associated with the described invention recited in the claims.
The preceding descriptions of functions, according to the present invention for ordering access to shared resources to threads within Primary and Backup replicas, has been embodied within functions contained in a Consistent Multithreading library that preferably intercepts calls to the operating system's thread library; however, it should be appreciated that the functions of the present invention may be incorporated, alternatively, into the operating system's thread library, or be otherwise configured without departing from the teachings of the present invention.
The description above has described functions and mechanisms that provide virtual determinism and strong replica consistency when multiple multithreaded application replicas are being executed. It should be appreciated that these functions and mechanisms may be modified in a number of different ways, and/or be implemented with reduced functionality, or added functionality, by one of ordinary skill in the art without departing from the present invention.
Queue based ordering and granting of mutex claims is described by way of example as a mechanism for ordering claims and suspending threads awaiting a mutex; however, it should be appreciated that any other convenient locking, tagging, or similar ordering means can be utilized, in combination with communication of such claims, by one of ordinary skill without departing from the teachings of the present invention.
The ordering and granting of claims to access shared resources by Backup replicas is described by way of example to use the delivery order of multicast messages from a Primary replica. However, other mechanisms may be utilized for communicating ordering information across a number of replicas as may be implemented by one of ordinary skill in the art without departing from the teachings of the present invention.
Multicast messages are exemplified containing thread, mutex and claim number information (T, M, N), wherein mutex claims may be distinguished from one another; however, it should be appreciated that claims may be distinguished from one another using alternative mechanisms and/or data structures, without departing from the teachings of the present invention. In a similar manner, the entries in a claims queue for a particular mutex are exemplified as containing fields for thread and claim number (T, N), wherein alternative means for distinguishing one claim from another may be adopted without departing from the teachings of the present invention. In the above situations, it should also be appreciated that fields may be added to the above claim descriptions to enhance functionality, or to accommodate additional enhancements and/or modifications to the described systems, software mechanisms, methods, improvements and apparatus, without departing from the present invention.
The use of multicast messages with piggybacked ordering information has been described for general use; however, messages may be delivered utilizing any convenient mechanism, such as message multicasting of a separate message (i.e. control message) containing the ordering information under select situations, or point-to-point messages, to communicate the ordering and granting of claims to multiple replicas.
The invention may be described as addressing the problem of maintaining strong replica consistency of replicated multithreaded application programs that employ the leader-follower (primary-backup) strategy of semi-active or passive replication, by masking multithreading as a source for non-determinism, and thus rendering a replicated multithreaded application program virtually deterministic, in a manner that is transparent to the application program.
To provide virtual determinism and to maintain strong replica consistency, the Primary replica determines the order in which accesses to shared resources are granted to threads in the Backup replicas, which is the same order as the corresponding accesses were granted to the corresponding threads in the Primary replica. In the Primary replica, when a request to access a shared resource is ordered and granted to a thread, the Primary replica multicasts a message containing the ordering and granting information to the Backup replicas. In a Backup replica, when a thread requests access to a shared resource, and when the multicast protocol orders and delivers a message containing the access request information from the Primary replica, and when the shared resource is available, the request is granted to the thread in the Backup replica.
To maintain transparency to the application program, the invention intercepts the calls to functions of the operating system's thread library and diverts them to a Consistent Multithreading library that is interposed ahead of the operating system's thread library. The Consistent Multithreading library contains wrapper functions of the functions of the operating system's thread library that claim and release mutexes, semaphores, condition variables, and so forth. When the Primary or Backup replicas invoke functions to claim or release shared resources, they invoke the wrapper functions of the Consistent Multithreading library, which sanitize the non-determinism due to multithreading that would otherwise exist and, in turn, invoke the corresponding functions of the operating system's thread library. Thus, the invention ensures that the Primary and Backup replicas make the same decisions when those functions are invoked, in a manner that is transparent to the application program.
Although the description above contains many specificities, these should not be construed as limiting the scope of the invention but as merely providing illustrations of some of the presently preferred embodiments of this invention. Thus, the scope of this invention should be determined by the appended claims and their legal equivalents. Therefore, it will be appreciated that the scope of the present invention fully encompasses other embodiments which may become obvious to those skilled in the art, and that the scope of the present invention is accordingly to be limited by nothing other than the appended claims, in which reference to an element in the singular is not intended to mean “one and only one” unless explicitly so stated, but rather “one or more.” All structural and functional equivalents to the elements of the above-described preferred embodiment that are known to those of ordinary skill in the art are expressly incorporated herein by reference and are intended to be encompassed by the present claims. Moreover, it is not necessary for a device or method to address each and every problem that the present invention seeks to solve, for it to be encompassed by the present claims. Furthermore, no element, component, or method step in the present disclosure is intended to be dedicated to the public regardless of whether the element, component, or method step is explicitly recited in the claims. No claim element herein is to be construed under the provisions of 35 U.S.C. 112, sixth paragraph, unless the element is expressly recited using the phrase “means for.”
This application claims priority from U.S. provisional application Ser. No. 60/367,615 filed on Mar. 25, 2002, incorporated herein by reference, and from U.S. provisional application Ser. No. 60/367,616 filed on Mar. 25, 2002, incorporated herein by reference.
This invention was made with Government support under Grant No. 70NANBOH3015, awarded by the U.S. Department of Commerce, National Institute of Standards and Technology. The Government may have certain rights in this invention.
Number | Name | Date | Kind |
---|---|---|---|
4718002 | Carr | Jan 1988 | A |
5257381 | Cook | Oct 1993 | A |
5577261 | Harinarayan et al. | Nov 1996 | A |
5621885 | Del Vigna, Jr. | Apr 1997 | A |
5794034 | Harinarayan et al. | Aug 1998 | A |
5799146 | Badovinatz et al. | Aug 1998 | A |
5802265 | Bressoud et al. | Sep 1998 | A |
5941999 | Matena et al. | Aug 1999 | A |
5956489 | San Andres et al. | Sep 1999 | A |
5968185 | Bressoud et al. | Oct 1999 | A |
6332200 | Meth et al. | Dec 2001 | B1 |
6463454 | Lumelsky et al. | Oct 2002 | B1 |
6539446 | Chan | Mar 2003 | B1 |
6574750 | Felber et al. | Jun 2003 | B1 |
6928577 | Moser et al. | Aug 2005 | B2 |
20020032883 | Kampe et al. | Mar 2002 | A1 |
20020099973 | Moser et al. | Jul 2002 | A1 |
20050229035 | Peleska et al. | Oct 2005 | A1 |
Number | Date | Country | |
---|---|---|---|
20040078618 A1 | Apr 2004 | US |
Number | Date | Country | |
---|---|---|---|
60367615 | Mar 2002 | US | |
60367616 | Mar 2002 | US |