R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile Logging in n-Fault-Tolerant Distributed Systems. In Proceedings of the 18th Annual International Symposium on Fault-Tolerant Computing, pages 44--49, 1988.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....state is inconsistent with the recovered state of a crashed process. Thus, message logging protocols guarantee either through careful logging or through a somewhat complex recovery protocol that after recovery no process is an orphan. Message logging protocols can be pessimistic (for example, [5, 11, 17, 24]) optimistic (for example, 12, 22, 23, 26] or causal [4] Like pessimistic protocols, causal protocols [3, 10] never create orphans, and, like optimistic protocols, they do not log synchronously to stable storage. They are able to do this by piggybacking information onto the ambient message ....

....its execution up to s p were sentby q during its execution up to s q , and vice versa. A collection of states, one from each process, is a consistent global state if all pairs of states are mutually consistent [6]# otherwise it is inconsistent. We assume that processes are piecewise deterministic [24] in that the only nondeterminism in a process arises from the nondeterministic order in which messages that have been received are delivered. It is therefore natural to think of the execution of a process as being partitioned into intervals, with the beginning of eachinterval being defined by the ....

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile Logging in n-Fault-Tolerant Distributed Systems. In Proceedings of the 18th Annual International Symposium on Fault-Tolerant Computing, pages 44--49, 1988.

....event that represents the delivery of a received message to the application or applications running in that process. For any message m from process p to process q, q delivers m only if it has received m, and q delivers m no more than once. We assume that processes are piecewise deterministic [7, 14], i.e. that it is possible to identify all the non deterministic events executed by each process and to log for each such event a determinant [4] that contains all the information necessary to replay the event during recovery. In particular, we assume that the order in which messages are delivered ....

.... n described in [3] and to Manetho [8] We then discuss its limitations with respect to scaling, and present a hierarchical and scalable causal message logging protocol. 3. 1 Review of SCML Like other message logging protocols, causal message logging is built using a recovery unit abstraction [14]. The recovery unit acts like a filter between the application and the transport layer. When an application sends a message, the recovery unit records fault tolerance information on the message and hands it off to the transport layer. Similarly, on the receiving end, the recovery unit reads the ....

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile logging in n-fault-tolerant distributed systems. In Proceedings of the Eighteenth Annual International Symposium on Fault-Tolerant Computing, pages 44-49, June 1988.

....As we noted above, the main problem in message logging derives from the difficulty of satisfying the consistency condition that must hold upon recovery of a faulty process in a way that is both simple and efficient. Yet, in spite of the numerous message logging protocols [SY85,JZ87,JV87,SBY88,SW89,JZ90,Eln93,VJ94] no precise specification of what such consistency condition requires has been presented in the literature. In this dissertation, we present the first formal specification of a necessary and sufficient condition for avoiding orphan processes. After showing how pessimistic ....

....from that of [CL85] in that it is defined in terms of deliver events rather than receive events. Our usage corresponds to the literature on message logging protocols. The transitive closure of OE over states yields the happens before relation over states. Processes are piecewise deterministic [SBY88] execution of a process consists of a sequence of deterministic intervals of execution, joined by non deterministic events. For each process, the first interval of execution begins with the process initial state; subsequent intervals begin with each nondeterministic event. Only deliver events ....

[Article contains additional citation context not shown here]

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile logging in n-fault- tolerant distributed systems. In Proceedings of the Eighteenth Annual International Symposium on Fault-Tolerant Computing, pages 44--49, 1988.

....of message logging, message logging protocols must guarantee that there are no orphan processes, either through careful logging or through a somewhat complex recovery protocol. The two main approaches to message logging are optimistic (for example, 18, 17, 11, 20] and pessimistic (for example, [7, 15, 10, 19]) We have recently defined a third approach that we call causal [3] There are two published causal message logging protocols: Family Based Logging (FBL) 4] and Manetho [9] In the same paper we defined a message logging protocol to be optimal if it is causal and does not send any additional ....

....execution up to s p were sent by q during its execution up to s q , and vice versa. A collection of states, one from each process, is a consistent global state if all pairs of states are mutually consistent [8] otherwise it is inconsistent. We assume that processes are piecewise deterministic [19] in that the only nondeterminism in a process arises from the nondeterministic order in which messages are delivered. It is therefore natural to think of the execution of a process as being partitioned into intervals, with the beginning of each interval being defined by the initial state of the ....

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile logging in n-fault-tolerant distributed systems. In Proceedings of the Eighteenth Annual International Symposium on Fault-Tolerant Computing, pages 44--49, 1988.

....event that represents the delivery of a received message to the application or applications running in that process. For any message m from process p to process q, q delivers m only if it has received m, and q delivers m no more than once. We assume that processes are piecewise deterministic [7, 14], i.e. that it is possible to identify all the non deterministic events executed by each process and to log for each such event a determinant [4] that contains all the information necessary to replay the event during recovery. In particular, we assume that the order in which messages are delivered ....

....f = n described in [3] and to Manetho [8] We then discuss its limitations with respect to scaling, and present a hierarchical and scalable causal message logging protocol. 3. 1 Review of SCML Like other message logging protocols, causal message logging is built using a recovery unit abstraction [14]. The recovery unit acts like a lter between the application and the transport layer. When an application sends a message, the recovery unit records fault tolerance information on the message and hands it o to the transport layer. Similarly, on the receiving end, the recovery unit reads the ....

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile logging in n-fault-tolerant distributed systems. In Proceedings of the Eighteenth Annual International Symposium on Fault-Tolerant Computing, pages 44-49, June 1988.

....in part by the Office of Naval Research under contract N00014 97 1 1013. collect this information [23, 28, 8, 21, 1, 20] Other positive aspects of coordinated checkpointing are that concurrent failures can be tolerated, and applications do not have to execute in a piece wise deterministic manner [22]. Arguments commonly used against coordinated checkpointing have been the overhead and lack of independence of checkpoint creation. Traditionally, two types of coordination have been employed: extra message exchanges and message tagging. A coordinated protocol sends extra messages, for example, ....

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile logging in n-fault-tolerant distributed systems. Proceedings of the 18th International Symposium on Fault-Tolerant Computing, pages 44--49, June 1988.

....into two classes: pessimistic and optimistic. Optimistic message logging protocols (for example, SY85, SW89, JZ90, VJ94] allow orphan processes to be created but eventually roll back their state to restore consistency. Pessimistic message logging protocols (for example, BBG83, PP83, JZ87, SBY88] never allow orphans to be created, but can introduce blocking in failure free runs. Recently, two message logging protocols that are neither optimistic nor pessimistic but combine the desirable properties of the two classical approaches have been proposed. These protocols, Family Based Logging ....

....up to s p were sent by q during its execution up to s q , and vice versa. A collection of states, one from each process, is a consistent global state if all pairs of states are mutually consistent [CL85] otherwise it is inconsistent. We assume that processes are piecewise deterministic [SBY88] in that the only nondeterminism in a process arises from the nondeterministic order in which messages are received. It is therefore natural to think of the execution of a process as being partitioned into intervals, with the beginning of each interval being defined by the initial state of the ....

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile logging in n-fault-tolerant distributed systems. In Proceedings of the Eighteenth Annual International Symposium on Fault-Tolerant Computing, pages 44--49, 1988.

....as program counters) to their current values. We assume that the state of the process does not include the state of the underlying communication system, such as the queue of messages that have been received but not yet delivered to the process. Execution of a process is piecewise deterministic [26]: It consists of a sequence of deterministic intervals of execution, joined by non deterministic events. For each process, the first interval of execution begins with the process initial state; subsequent intervals begin with each non deterministic event. Hence, execution of a process consists of ....

....a crashed process. Consistency for log based protocols translates into the guarantee that upon recovery no process is an orphan. 4 The performance of a logging based protocol depends heavily on how the protocol enforces the noorphans guarantee. For instance, pessimistic protocols (for example, [7, 19, 13, 26]) never create orphans by logging determinants on stable storage synchronously. Unfortunately, these protocols exhibit relatively poor performance during failure free runs, since they prevent processes from communicating until logging is complete. In contrast, optimistic protocols protocols [25, ....

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile Logging in n-Fault-Tolerant Distributed Systems. In Proceedings of the Eighteenth Annual International Symposium on Fault-Tolerant Computing, pages 44--49, 1988.

....a simple, uniform approach, which can provide low overhead fault tolerance to applications in which communication is performed through message passing, file sharing, or a combination of the two. 1 Introduction Low overhead rollback recovery protocols such as checkpointing and message logging [2, 3, 9, 17, 18] have been extensively studied for message passing applications. These protocols seek to tolerate common failures while minimizing the use of additional resources and the impact on performance during failure free executions. In this paper, we focus on low overhead protocols for applications in ....

....that p is part of. Deliver, read and write events are non deterministic, because the order in which an agent receives messages and the file versions it accesses are execution dependent. Send events and other local events are instead deterministic. Agent execution is piecewise deterministic [17]: It consists of a sequence of deterministic intervals of execution, joined by non deterministic events. At any point during the execution, the state of an agent is a mapping of program variables and implicit variables (such as program counters) to their current values 1 . Given the initial ....

[Article contains additional citation context not shown here]

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile Logging in n-Fault-Tolerant Distributed Systems. In Proceedings of the Eighteenth Annual International Symposium on Fault-Tolerant Computing, pages 44-- 49, 1988.

....multiple host failures. We note in passing that similar functionality 27 has been implemented elsewhere by logging messages onto non volatile storage as they are received [24] logging them at a monitor site [22] or by retaining copies of messages in the volatile storage of sending processes [15,16,25]. There are two conditions that must be satisfied to guarantee recreation of the appropriate state and view. One is that each participant in a conversation must be deterministic. For our purposes, this means that a process s state transitions and generated messages (i.e. its output) are ....

.... for group communication [8] the exchange of very large messages [9,29] alternative send receive semantics [6] guaranteeing a consistent order on message delivery in a manyto many communication [3,4] and techniques for logging messages so as to facilitate recovery from processor failure [15,16,22,24,25]. The work presented in this paper addresses the latter two issues. Psync is most closely related to the ISIS protocol suite ABCAST (atomic broadcast) CBCAST (causal broadcast) and GBCAST (group broadcast) 3,4] From Psync s perspective, ABCAST and CBCAST are specific message ordering ....

R. Strom, D. Bacon, and S.Yemini. Volatile logging in n-fault-tolerant distributed systems. In Proceedings of the Eighteenth International Symposium on Fault-Tolerant Computing (June 1988), to appear.

....lessons that can guide the design of message logging protocols that combine low overhead during failure free executions with fast recovery. Technical Report TR 98 02, Department of Computer Sciences, University of Texas, Austin, TX. 1 1 Introduction Message logging protocols (for example, [1, 4, 5, 9, 12, 13, 14, 18, 20, 21, 22, 23]) are popular techniques for building systems that can tolerate process crash failures. These protocols are built on the assumption that the state of a process is determined by its initial state and by the sequence of messages it delivers. In principle, a crashed process can be recovered by (1) ....

.... The second protocol is instead sender based [12] the receiver logs synchronously to stable storage only the determinant of every message it delivers, while the contents of the message are stored in a volatile log kept by the message s sender 2 This protocol is similar to the one described in [22]. In both of these protocols, the first step of recovering a process p consists in restoring it to its latest checkpoint. Then, in the receiver based protocol, the messages logged on stable storage are replayed to p in the appropriate order. In the sender based protocol, instead, p broadcasts a ....

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile Logging in n-Fault-Tolerant Distributed Systems. In Proceedings of the Eighteenth Annual International Symposium on Fault-Tolerant Computing, pages 44--49, 1988.

....their determinant, the implementation 5 of stable storage, etc. Figure 2 shows a set of representative instantiations of these variables used in several existing protocols. Figure 3 illustrates how the language can be used to specify Bacon, Strom, and Yemini s sender based pessimistic protocol [33], Damani and Garg s optimistic protocol [7] and a hybrid protocol we have recently developed that combines causal logging with asynchronous receiver based logging to stable storage in order to speed up crash recovery [26] Our language, however, can be used for more than just specifying existing ....

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile Logging in n-Fault-Tolerant Distributed Systems. In Proceedings of the Eighteenth Annual International Symposium on Fault-Tolerant Computing, pages 44--49, 1988.

....are no failures. This is the main disadvantage of pessimistic message logging. 5.2. 2 Optimistic message logging In the optimistic (also called asynchronous) message logging approach, messages are processed by application computation independently of and concurrently to their stable logging ([4, 10, 13, 15, 16, 17, 27, 37, 39, 43, 44]) Sent or received messages are, first, logged in volatile storage with negligeable interference with respect to application computations. Then periodically, or when an application process is idle, its controller independently saves volatile logs in stable storage to make them stable, and clears ....

R.E. Strom, D.F. Bacon, S.A. Yemini, Volatile logging in n-fault-tolerant distributed systems, Proc. Fault Tolerant Computing Systems, 1988, pp. 44-49.

....tolerance using checkpointing. The Manetho system [24] did not provide predicate detection or optimism. In addition we show in Chapter 4 that our recovery scheme allows us to send output commit messages faster than the approach of Manetho. Other checkpointing approaches also suffer the same delay [2, 15, 24, 40, 41, 64, 59, 61, 65]. 8 Our work is designed for distributed computations made up of processes communicating via messages, in which processes may crash and messages may not arrive at their destinations. We have designed algorithms for each of the four areas discussed in Section 1.3. Each process in the computation ....

....ordering. No coordination of checkpointing is required in these schemes. Pessimistic message logging [10, 54] forces a process to wait before sending any message while the message log is written to stable storage. Optimistic logging methods [41, 59, 61, 65] and the similar sender based logging [40, 64]) assume failures are rare and therefore allow ordering information to be lost in a failure. That is, a message is logged in the background while execution proceeds) Consequently, received messages and any sends that depend on them may not be recoverable. This may then require that unfailed ....

R. Strom, D. Bacon, and S. Yemeni. Volatile logging in n-fault-tolerant distributed systems. Proceedings fo the 18th International Symposium on FaultTolerant Computing, pages 44--49, June 1988.

....Thus, message logging protocols must guarantee that there are no orphan processes after recovery either through careful logging or through a somewhat complex recovery protocol. The three approaches to message logging are optimistic (for example, 13, 20, 21, 23] pessimistic (for example, [6, 12, 17, 22]) and causal [2] Like pessimistic protocols, causal protocols [4, 11] never create orphans, and like optimistic protocols, they do not synchronously log to stable storage. They are able to do this by piggybacking information onto the ambient message traffic. Causal message logging protocols ....

....execution up to s p were sent by q during its execution up to s q , and vice versa. A collection of states, one from each process, is a consistent global state if all pairs of states are mutually consistent [7] otherwise it is inconsistent. We assume that processes are piecewise deterministic [22] in that the only nondeterminism in a process arises from the nondeterministic order in which messages that have been received are delivered. It is therefore natural to think of the execution of a process as being partitioned into intervals, with the beginning of each interval being defined by the ....

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile logging in n-fault-tolerant distributed systems. In Proceedings of the Eighteenth Annual International Symposium on Fault-Tolerant Computing, pages 44--49, 1988.

....Computing and Systems (ICDCS 98) Amsterdam, Netherlands, May 1998 (to appear) Also available as Technical Report TR 98 01, Department of Computer Sciences, University of Texas, Austin, TX. 1 Introduction Low overhead fault tolerance protocols such as checkpointing and message logging [2, 3, 4, 13, 14, 17, 21, 27, 32, 30] have been extensively studied for message passing distributed applications. These protocols seek to tolerate common failures while minimizing the use of additional resources and the impact on performance during failure free executions. In this paper, we focus on low overhead protocols for ....

....that p is part of. Deliver, read and write events are non deterministic, because the order in which an agent receives messages and the file versions it accesses are execution dependent. Send events and other local events are instead deterministic. Agent execution is piecewise deterministic [27]: It consists of a sequence of deterministic intervals of execution, joined by non deterministic events. At any point during the execution, the state of an agent is a mapping of program variables and implicit variables (such as program counters) to their current values 1 . Given the initial ....

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile Logging in n-Fault-Tolerant Distributed Systems. In Proceedings of the Eighteenth Annual International Symposium on Fault-Tolerant Computing, pages 44--49, 1988.

....tradeoff, we propose hybrid logging protocols, a new class of orphan free protocols. We show that hybrid protocols perform within 2 of causal logging during failure free execution, and within 2 of receiver based logging during recovery. 1 Introduction Message logging protocols (for example, [2, 3, 6, 8, 11, 12, 19, 20]) are popular techniques for building systems that can tolerate process crash failures. These protocols are built on the assumption that the state of a process is determined by its initial state and by the sequence of messages it delivers. In principle, a crashed process can be recovered by (1) ....

....second protocol is instead sender based [11] the receiver logs synchronously to stable storage only the determinant of every message it delivers, while the contents of the message are stored in a volatile send log kept by the message s sender 2 . This protocol is similar to the one described in [20]. In both of these protocols, the first step of recovering a process p consists in restoring it to its latest checkpoint. Then, in the receiver based protocol, the messages logged on stable storage are replayed to p in the 2 Some sender based pessimistic protocols keep both determinants and ....

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile Logging in n-Fault-Tolerant Distributed Systems. In Proceedings of the Eighteenth Annual International Symposium on Fault-Tolerant Computing, pages 44--49, 1988.

....are optimal with respect to these metrics. Finally, starting from a simple protocol that relies on causal delivery order, we show how to derive optimal causal protocols that tolerate f overlapping failures and recoveries for a parameter f : 1 f n. 1 Introduction Message logging (for example, [BBG83, PP83, SY85, JZ87, SBY88, SW89, JZ90, EZ92]) is a common technique used to build systems that can tolerate process crash failures. These protocols require that each process periodically record its local state and log the messages it received after having recorded that state. When a process crashes, a new process is created in its place: ....

....up to s p were sent by q during its execution up to s q , and vice versa. A collection of states, one from each process, is a consistent global state if all pairs of states are mutually consistent [CL85] 1 otherwise it is inconsistent. We assume that processes are piecewise deterministic [SBY88] in that the only nondeterminism in a process arises from the nondeterministic order in which messages are received. It is therefore natural to think of the execution of a process as being partitioned into intervals, with the beginning of each interval being defined by the initial state of the ....

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile logging in n-fault-tolerant distributed systems. In Proceedings of the Eighteenth Annual International Symposium on Fault-Tolerant Computing, pages 44--49, 1988.

....the current costs of high availability, there is a fresh need for recovery based techniques that combine high performance during failurefree executions with fast recovery. As an initial step towards the development of these new techniques, we have implemented a sender based pessimistic protocol [5], a receiver based pessimistic protocol [2] and a causal protocol [1] and studied their performance during recovery [4] 1 . All of these protocols log (1) the content and (2) the order of receipt of each message delivered by each process during its execution. Processes synchronously log on ....

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile logging in n-fault-tolerant distributed systems. In Proceedings of the Eighteenth Annual International Symposium on FaultTolerant Computing, pages 44--49, 1988. Application f Restore Acquire Roll Total Percentage

....correct processes to roll back together, in order to resolve the inconsistent global state. In sender based logging protocols, the messages for recovery are saved in the sender s volatile storage, while only the receipt sequence numbers of those messages are logged in the receiver s stable storage[7, 11]. Causal message logging protocols combine the advantages of all classes by keeping each message in the volatile log of its sender and saving its receipt sequence number in the volatile log of causally dependent processes[5, 1] Hence, they do not require communication induced blocking for stable ....

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile logging in n-fault-tolerant distributed systems. In Digest of Papers: Annual International Symposium on Fault-Tolerant Computing, pages 44--49, 1988.

.... consider that processes execute programs which can have non deterministic statements (i.e. processes histories cannot be entirely predicted by programs these processes execute) Let us note that this process s behavior includes, but is not limited to, the piecewise deterministic one presented in [13] where the only cause of non determinism is due to the unpredictability of message transmission delays. RR n2569 6 Roberto Baldoni, Jean Michel Helary, Achour Mostefaoui, Michel Raynal 1 C 0 0 i i 2 i 5 i 6 i 7 C 2 i i 8 i 4 C 1 i 3 0 S 0 k 1 0 2 2 0 C C 1 j j P P Local State Checkpoint Interval ....

R.E. Strom, D.F. Bacon, S.A. Yemini, Volatile logging in n-fault-tolerant distributed systems, Proc. 18-th International Conference of Fault Tolerant Computing Systems, 1988, pp. 44-49.

....for applications executing in a distributed system. With rollback recovery, information saved on stable storage during failure free execution allows certain states of each process to be recovered after a failure. Examples of such methods include the use of message logging and checkpointing [12, 8, 3, 15, 14, 9, 13], and the use of checkpointing alone [11, 4, 2] Optimistic methods in general allow unrecoverable states of one process to be seen by other processes, and optimistically assume that these states will become recoverable before a failure occurs. This allows the needed recovery information to be ....

Robert E. Strom, David F. Bacon, and Shaula A. Yemini. Volatile logging in n-fault-tolerant distributed systems. In The Eighteenth Annual International Symposium on Fault-Tolerant Computing: Digest of Papers, pages 44--49. IEEE Computer Society, June 1988.

....respect to these metrics. We give several examples of optimal message logging protocols that can tolerate f overlapping failures and recoveries for a parameter f : 1 f n, and discuss the tradeoffs that arise in the implementation of these protocols. 1 Introduction Message logging (for example, [BBG83, PP83, SY85, JZ87, SBY88, SW89, JZ90, EZ92]) is a common technique used to build systems that can tolerate process crash failures. These protocols require that each process periodically record its local state and log the messages it received after having recorded that state. When a process crashes, a new process is created in its place: ....

....up to s p were sent by q during its execution up to s q , and vice versa. A collection of states, one from each process, is a consistent global state if all pairs of states are mutually consistent [CL85] 1 otherwise it is inconsistent. We assume that processes are piecewise deterministic [SBY88] in that the only nondeterminism in a process arises from the nondeterministic order in which messages are received. It is therefore natural to think of the execution of a process as being partitioned into intervals, with the beginning of each interval being defined by the initial state of the ....

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile logging in n-fault-tolerant distributed systems. In Proceedings of the Eighteenth Annual International Symposium on Fault-Tolerant Computing, pages 44--49, 1988.

....low overhead solution for building distributed applications that tolerate crash failures. These protocols come in several flavors. Checkpoints can be independent, coordinated, or induced by specific patterns of communication. Logging can be pessimistic, optimistic, or causal. Pessimistic protocols [3, 28] allow processes to communicate only from recoverable states. These protocols enforce this condition by synchronously logging to stable storage any information critical for recovery before letting processes communicate. Optimistic protocols (for instance, 7, 14, 27] allow processes to ....

....a non deterministic event broadcasts the corresponding determinant to the other processes and does not send any application message until it has received at least f acknowledgments to its broadcast. This protocol is a variant of Strom, Bacon, and Yemini s sender based pessimistic protocol [28], whose specification is shown in Figure 4(a) 4 The Architecture of Egida Figure 1 defines the structure of log based rollback recovery protocols, while the variables in Figure 2 identify the building blocks which when incorporated into the protocol structure yield different rollback recovery ....

R. E. Strom, D. F. Bacon, and S. A. Yemini. Volatile Logging in n-Fault-Tolerant Distributed Systems. In Proceedings of the Eighteenth Annual International Symposium on Fault-Tolerant Computing, pages 44--49, 1988.

....memory virtualizes the space of a process, we virtualize time, paging in a previous process state when a time fault (or incorrect guess) occurs. 2 Optimistic Recovery Our approach to recovery is based upon optimistic recovery [6] enhanced by optimizations to reduce the amount of logging [5, 1] and extensions which incorporate the filesystem and other external components into the recovery process [7, 8] In optimistic recovery, we guess that processors do not fail; specifically, for every non deterministic event (usually a message) we guess that there will not be a failure before that ....

STROM, R. E., BACON, D. F., AND YEMINI, S. A. Volatile logging in n-fault-tolerant distributed systems. In The Eighteenth Annual Interna- tional Symposium on Fault-Tolerant Computing: Digest of Papers (June 1988), pp. 44-49.

First 50 documents

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC