L. Alvisi. Understanding the message logging paradigm for masking process crashes. PhD thesis, Cornell University, Department of Computer Science, 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....t 2 S : t 2 run s t ) s 2 run 2. Recovering after failures When a process fails in causal message logging algorithms, to recover it must recreate its message log (that is, the message receive ordering) The information necessary to encapsulate the ordering of messages is called determinants [1, 2]. We will use this concept here. A recovery algorithm for causal message logging must ensure that messages from recovering processes do not create orphan states. Figure 2 illustrates how this might happen when recovering a single process. In the figure, P 1 , upon restart, requests all ....

....m m m recovery request order of not recovered While it may seem that the solution is to have m 00 include the incarnation of P 1 as well as P 2 , there is a less costly solution, which we will present in the next section. First, however, we will discuss previous solutions. Alvisi, [1], and Johnson and Zwaenepoel [9] avoid this problem using the following idea: processes that respond with determinant information are required to first write this information to stable storage. These processes must then also block the receive of any messages until after a special message has been ....

[Article contains additional citation context not shown here]

L. Alvisi. "Understanding the message logging paradigm for masking process crashes." In Ph.D. dissertation. Cornell University, January 1996.

....to stable storage (storage for which data persists even in the event of failures) Because processes are not replicated, when a failure occurs, the saved checkpoint must be reloaded so that the entire computation does not have to restart from the beginning. The different checkpointing methods [2, 7, 10, 20, 26, 24, 41, 44, 54, 59, 61, 65] offer other tradeoffs which we will discuss in Section 1.3.3. The Manetho system employs both active replication and checkpointing [24] Active replication is used for system processes (for example, name services) and a checkpointing scheme is designed for applications. As mentioned above, the ....

....earlier than completion of recovery. ffl How can the latencies associated with recovery be decreased Another area to consider for reducing delay of the application is the process recovery method itself. In addition to the time it takes to recover a failed process, some checkpointing methods [2, 41] can cause delays when sending and receiving certain messages (for example, messages received during recovery) Other methods [7, 20, 26, 41, 44, 59, 61, 65] can cause unfailed processes to go back to the state of a previous checkpoint when a failure occurs. Finally, checkpointing schemes impose ....

[Article contains additional citation context not shown here]

L. Alvisi. Understanding the message logging paradigm for masking process crashes. In Ph.D. dissertation. Cornell University, January 1996.

....should it be sparse. The resulting version of Pi jLogj piggybacks no more than f additional words than Pi Det , an amount which is independentofD. A drawback of this approach, however, is that determinants sorted in this manner are not suitable for some of the compression techniques described in [2, 3], which can dramatically reduce the size of the piggyback. Furthermore, while this this approach can also be applied to Pi , it can not be applied to Pi Log or Pi . In the next section we introduce a data structure, called a dependency matrix, that allows us to implement Pi with ....

L. Alvisi. Understanding the Message Logging Paradigm for Masking Process Crashes. PhD thesis, Cornell University Department of Computer Science, January 1996.

....most efficient even when f = n. Due to lack of space, we do not present the protocol that is run when a crashed process recovers. All five protocols that we develop in this paper can use the same recovery protocol. A discussion on recovery as well as the actual recovery protocol can be found in [1]. 2 System Model We assume a system N of n processes that can communicate only by exchanging messages. The system is asynchronous: there exists no bound on the relative speeds of processes, no bound on message transmission delays, and no global time source. The execution of the system is ....

....and edges are used to represent the direction application messages are sent. Definition 1 A channel graph is tree like if, for all pairs of nodes i and j, all paths from i to j have the same length. Note that the channel graph of Figure 3 is tree like. The following theorem, proved in [1], characterizes one set of applications for which Pi Det performs as well as Log when f = n. Theorem 1 Let f = n. Given a tree like channel graph, for any run ae, Protocol Pi Det piggybacks on each message the same determinants as Protocol Pi Log . Figure 2: Comparison of Pi Log and ....

[Article contains additional citation context not shown here]

L. Alvisi. Understanding the Message Logging Paradigm for Masking Process Crashes. PhD thesis, Cornell University Department of Computer Science, January 1996.

....more than f entries of D p [ m.dest] are greater than, or equal to, m.rsn. Because the order of events executed by a processor is in fact a total order, it is also straightforward to construct a dependency matrix that has size n P Theta n P where n P is the number of processors in the system [Alv95] When, in the following, we discuss the cost of the protocols, we assume that this smaller representation of the dependency matrix is used. 6.3 Piggybacking the Dependency Matrix Since the dependency matrix of process p can be used to compute Log(m) p for all the messages for which p is a ....

L. Alvisi. Understanding the Message Logging Paradigm for Masking Process Crashes. PhD thesis, Cornell University Department of Computer Science, expected completion 1995.

....this tradeoff, however, it is instructive to compare the amount of data that each protocol piggybacks on a message carrying a fixed number of determinants. For simplicity, in our calculations we don t consider optimizations achievable by applying compression techniques such as those described in [1, 4]. Consider a message m from process p to process q. Suppose that, when p sends m, UnstableDL p (q) contains D determinants and DL p contains N determinants. Let w denote the number of words needed to encode a determinant, and assume that the identity of a process and the number of processes that ....

....jLog(m 0 )j. The resulting version of Pi jlog j piggybacks no more than f additional words than Pi det , an amount which is independent of D. A drawback of this approach, however, is that determinants sorted in this manner are not suitable for some of the compression techniques described in [1, 4], which can dramatically reduce the size of the piggyback. Furthermore, this approach can not be applied to Pi jlog j , Pi log or Pi log . In the next section we introduce a data structure, called a dependency matrix, that allows us to implement Pi det , Pi jlog j , and Pi ....

L. Alvisi. Understanding the Message Logging Paradigm for Masking Process Crashes. PhD thesis, Cornell University Department of Computer Science, January 1996.

....if the read event does not result in the acquisition of a new version of the file; and (2) by generating a determinant only when p acquires a new file version. We treat write events similarly. The resulting determinants can be efficiently encoded using compression techniques discussed elsewhere [1, 2]. We can also reduce the number of file versions logged in the volatile memory of agents. Suppose an agent p acquires write ownership on a file F and repeatedly modifies it before its ownership is revoked. Since all but the last version of F created by p are invisible to other agents in the ....

L. Alvisi. Understanding the Message Logging Paradigm for Masking Process Crashes. PhD thesis, Cornell University Department of Computer Science, January 1996.

....protocols cope with this loss of information. There are other issues of message logging protocols that are out of the scope of this paper, such as communication with the environment, checkpointing, and recovery. A discussion of these issues in the context of causal message logging is presented in [Alv96]. The paper proceeds as follows. Section 2 describes the system model commonly assumed for message logging protocols. Section 3 discusses the notion of consistency in message logging protocols. Section 4 presents the derivation of the always no orphans condition, and explains how it relates with ....

....f overlapping failures and recoveries for a parameter f : 1 f n by showing how such a protocol can be derived starting from a simple protocol that relies on causal delivery order. A thorough discussion of optimal protocols and of the tradeoffs involved in their implementation is presented in [AM96, Alv96]. Acknowledgments We would like to thank Bruce Hoppe, Fred Schneider and Sam Toueg for many discussions on this research and on early versions of this paper. We would also like to thank the anonymous referees for their helpful suggestions. A Derivations In this section, we show the derivations ....

[Article contains additional citation context not shown here]

L. Alvisi. Understanding the Message Logging Paradigm for Masking Process Crashes. PhD thesis, Cornell University Department of Computer Science, January 1996.

....jLog(m)j p is p s estimate of the size of Log(p) which can simply be the size of p s estimate Log p (m) of Log(m) Family based logging (FBL) is a family of causal message logging protocols that use different techniques for computing Log p (m) and jLog(m)j p . The five FBL protocols presented in [3, 4] differ in how accurately they compute Log p (m) and jLog(m)j p . They also differ in the amount of information they piggyback on messages: roughly speaking, the more information that is piggybacked, the more accurate the estimates are. It is not obvious, however, which of these five protocols ....

....protocol P 5 extends P 3 by piggybacking a dependency matrix of size n Theta n. This matrix is a vector of vector clocks [13] based on receive sequence numbers, and therefore encodes exactly Log(m) restricted to the causal past of p. Further details of protcols P 1 through P 5 can be found in [3, 4]. Table 1 summarizes the piggybacking overhead of the six family based logging protocols. In this table, W is the number of bits needed to represent a word that can hold an entry in the stability matrix, a component of a vector clock, a send sequence number, and so on. Manetho [7] is a variation ....

L. Alvisi. Understanding the Message Logging Paradigm for Masking Process Crashes. PhD thesis, Cornell University Department of Computer Science, January 1996.

No context found.

L. Alvisi. Understanding the message logging paradigm for masking process crashes. PhD thesis, Cornell University, Department of Computer Science, 1996.

No context found.

L.Alvisi. Understanding the message logging paradigm for masking process crashes. PhD thesis, Cornell University, January 1996.

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