L. Sabel and K. Marzullo. Election vs. consensus in asynchronous systems. TR 95-1488, Computer Science Department, Cornell University, Ithaca, New York, February 1995.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....perfectly if a remote process is crashed, e.g. if one can use a local hardware clock to decide if a time quantum of a remote process has expired. A. Termination Vs Conditional Timeliness Conditions There is no commonly agreed upon rigorous specification for the election problem. For example, [29] specifies the election problem for the time free system model as follows: TF) infinitely often there exists a leader, i.e. for any real time there exists a real time and a process becomes leader at . Typically, problems specified for timed systems do not use such strong ....

....asynchronous systems as follows: TT) when a majority of processes are F stable in a time interval E GFd ; H , then there exists a process that becomes leader GFj ; H . The specification is not implementable in time free systems, even when only one process is allowed to crash [29], while F, is implementable in timed systems [10] 13] To explain why this is so, consider a time free system that contains at least two processes . To implement , one has to solve the following problem: when a process becomes leader and stays leader until it crashes at a ....

[Article contains additional citation context not shown here]

L. Sabel and K. Marzullo. Election vs. consensus in asynchronous systems. Technical Report TR95-1488, Cornell University, Feb 1995.

....before it has crashed [3] A Perfect failure detector satisfies strong completeness and strong accuracy. The (highly available leader) election problem [4] requires that at any point in real time there be at most one leader and for any time for which there is a leader. A recent article [9] proves that a Perfect failure detector is the weakest reasonable failure detector which allows a deterministic solution of the election problem, where reasonable means a failure detector that satisfies certain symmetry conditions. Our experience indicates that (1) a Perfect failure detector ....

....introduced accuracy and completeness properties [3] with a fail aware property to define new fail aware failure detector classes. Sabel and Marzullo prove that a Perfect failure detector is the weakest failure detector to solve the election problem in the class of reasonable failure detectors [9]. This class contains all failure detectors which satisfy some given symmetry conditions. In the newest version of their paper they explicitly exclude the case that a process can use its local failure detector module to derive that other processes might wrongly suspect . Fail aware failure ....

[Article contains additional citation context not shown here]

L. Sabel and K. Marzullo. Election vs. consensus in asynchronous systems. Technical Report TR95-1488, Cornell University, Feb 1995. 18

....solve NBAC in any environment f with f #. From [1] 6] and Corollary 6, we have: Corollary 7 For any environment f with 0 f in a system augmented with timeless failure detectors, NBAC is strictly harder than Consensus. 6. Concluding remarks Sabel and Marzullo showed in [8] that is the weakest failure detector to solve the Leader Election problem within a specific class of failure detectors. They focus on failure detectors that output sets of suspected processes and satisfy the following symmetry property: if a process detects a failure erroneously, then any ....

....all failure detectors that make a finite number of mistakes. The approach is somewhat similar to ours. We also defined a subset of the overall universe of failure detectors [2] in which P #S is shown to be the weakest to solve our NBAC problem. The class of symmetric failure detectors of [8] and our class of timeless failure detectors are however incomparable. Fromentin, Raynal and Tronel stated in [4] that is the weakest failure detector to solve NBAC. Guerraoui [6] pointed out that [4] assumes NBAC to be solved among any subset of the processes in the system and showed ....

L. S. Sabel and K. Marzullo. Election vs. consensus in asynchronous systems. Technical report, Cornell University, Ithaca, NY, TR95-1488, 1995.

.... to agree (or even eventually agree) Introduced in [CT96] the abstraction of unreliable failure detectors has been used to solve several important problems such as consensus, atomic broadcast, group membership, non blocking atomic commitment, and leader election [BDM97, Gue95, HMR97, LH94, OGS97, SM95] Our goal is to use unreliable failure detectors to achieve quiescence, but before we do so we must address the following important question. Note that any reasonable implementation of a failure detector in a message passing system is itself not quiescent: a process being monitored by a failure ....

....can be used to solve this problem is 3P which is not implementable. 3. To overcome this obstacle, we introduce HB: this failure detector can be used to achieve quiescent reliable communication and it is implementable. In contrast to common failure detectors [BDM97, CT96, Gue95, GLS95, LH94, SM95] does not output a list of suspects, and it can be implemented without timeouts. 4. We show that can be used to extend existing algorithms for many fundamental problems (e.g. consensus, atomic broadcast, k set agreement, atomic commitment, approximate agreement) to tolerate message ....

Laura S. Sabel and Keith Marzullo. Election vs. consensus in asynchronous systems. Technical Report 95-1488, Department of Computer Science, Cornell University, Ithaca, New York, Febrary 1995.

....a fundamental problem in distributed systems, can be formulated as follows: all correct processes from a set of processes propose a value and then they have to decide on a common value. A distributed consensus server can provide a basic building block to implement other fault tolerant services [8, 3, 13, 9]. More formally, the consensus problem can be specified in terms of the following properties [12] Termination: Every correct process eventually decides some value. Validity: If a process decides v, then v was proposed by some process. Agreement: No two processes decide differently. In this ....

L. S. Sabel and K. Marzullo. Election vs. consensus in asynchronous systems. Technical Report TR95-1488, Cornel University, Feb. 1995.

....instances of TRB (multiple instances for each process as the distinguished sender) TRB is not the only natural problem that can be solved using P but cannot be solved using 3W. Other examples include the non blocking atomic commitment problem [CL94, Gue95] and a form of leader election [SM95] Figure 9 summarises these results. 9 Related work 9.1 Partial synchrony Fischer, Lynch and Paterson showed that Consensus cannot be solved in an asynchronous system subject to crash failures [FLP85] The fundamental reason why Consensus cannot be solved in completely asynchronous systems is ....

Laura Sabel and Keith Marzullo. Election vs. consensus in asynchronous systems. Technical Report TR95-411, University of California at San Diego, February 1995. Available at ftp://ftp.cs.cornell.edu/pub/sabel/tr94-1413.ps.

....execution environment called DSM Threads [14, 15] In this work, we assume asynchronous and reliable communication (no loss, duplication or modification of messages) with a bounded message latency. The bounded latency is required to counter the Consensus and the even harder election problem [23]. These problems are addressed in a practice by assuming that a node is faulty if it does not respond within the interval bounded by the latency. 2. Related Work A number of protocols for distributed synchronization have been published, in particular to address the problem of mutual exclusion in ....

L. S. Sabel and K. Marzullo. Election vs. consensus in asynchronous systems. Technical Report TR95-1488, Cornell University, Computer Science Department, Feb. 1995.

.... to agree (or even eventually agree) Introduced in [CT96] the abstraction of unreliable failure detectors has been used to solve several important problems such as consensus, atomic broadcast, group membership, non blocking atomic commitment, and leader election [BDM97, Gue95, HMR97, LH94, OGS97, SM95] Our goal is to use unreliable failure detectors to achieve quiescence, but before we do so we must address the following important question. Note that any reasonable implementation of a failure detector in a message passing system is itself not quiescent: a process being monitored by a failure ....

....can be used to solve this problem is 3P which is not implementable. 3. To overcome this obstacle, we introduce HB: this failure detector can be used to achieve quiescent reliable communication and it is implementable. In contrast to common failure detectors [BDM97, CT96, Gue95, GLS95, LH94, SM95] HB does not output a list of suspects, and it can be implemented without timeouts. 4. We show that HB can be used to extend existing algorithms for many fundamental problems (e.g. consensus, atomic broadcast, k set agreement, atomic commitment, approximate agreement) to tolerate message ....

Laura S. Sabel and Keith Marzullo. Election vs. consensus in asynchronous systems. Technical Report 95-1488, Department of Computer Science, Cornell University, Ithaca, New York, Febrary 1995.

....detector. It is unclear whether this specification will withstand the scrutiny of serious examination. 3.3 Leader Election Leader Election in fault tolerant distributed computing is an important problem for applications that provide replicated services. The Leader Election problem studied in [SM95] is specified as follows: Safety: At any time, at most one processor is the leader. Liveness: If there is no leader, a new leader will eventually be elected. Implicitly, the authors assume that if the leader crashes it is no longer the leader. The authors show that Leader Election can be ....

Laura Sabel and Keith Marzullo. Election vs. consensus in asynchronous systems. Technical Report CS95-411, Department of Computer Science, University of California, San Diego, February 1995.

....are correct and which are not. Distributed algorithms usually di#er on the assumptions made about the reliability of that mechanism. Some algorithms assume failure detectors that accurately detect crashes. For example, the state machine replication algorithm of [15] the election algorithm of [12], and the non blocking atomic commit algorithm of [16] assume that any correct process p i accurately detects when any other process p j has failed. Other algorithms do make weaker assumptions about failure detectors. For instance, none of the consensus algorithms of [2, 5, 9, 13] or the ....

L. Sabel and K. Marzullo. Election Vs. Consensus in Asynchronous Systems. Technical Report TR95-1488, Cornell Univ, 1995. Also, Technical Report CS95-411, UCSD, 1995.

.... the Hardness of Failure Sensitive Agreement Problems Rachid Guerraoui Swiss Federal Institute of Technology, CH 1015 Lausanne Abstract In [1] 3] and [6], respectively, it was stated that the weakest failure detector for any of non blocking atomic commit, terminating reliable broadcast and leader election, is the Perfect failure detector P. This paper presents a counter example of those results. We exhibit a failure detector that is ....

.... broadcast, the processes need to agree on whether to deliver a message broadcast by some specific sender process, or to deliver a default message [4] In leader election, the processes must elect a leader and make sure to avoid any disagreement about which process is leader at any given time [6]. Besides the fact that those three problems are all agreement problems, they also have a common failure sensitive flavor: in each of those problems, the decision value depends somehow on the failure pattern (i.e. on the fact that some processes 1 have crashed or not) 1 . In atomic commit, ....

[Article contains additional citation context not shown here]

L. Sabel and K. Marzullo. Election Vs. Consensus in Asynchronous Systems. Technical Report TR95-1488, Cornell Univ, 1995. Also, Technical Report CS95-411, UCSD, 1995. 9

....perfectly if a remote process is crashed, e.g. if one can use a local hardware clock to decide if a time quantum of a remote process has expired. A. Termination Vs Conditional Timeliness Conditions There is no commonly agreed upon rigorous specification for the election problem. For example, [29] specifies the election problem for the time free system model as follows: S) at any real time there exists at most one leader, and 14 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, TO APPEAR IN 1999 (TF) infinitely often there exists a leader, i.e. for any realtime s there exists a ....

.... exists at most one leader, and (TT) when a majority of processes are Delta F stable in a time interval [s; s ] then there exists a process p that becomes leader in [s; s ] The specification (S; TF ) is not implementable in time free systems, even when only one process is allowed to crash [29], while (S; TT ) is implementable in timed systems [10] 13] To explain why this is so, consider a time free system that contains at least two processes p and q. To implement (S; TF ) one has to solve the following problem: when a process p becomes leader at some real time s and stays leader ....

[Article contains additional citation context not shown here]

L. Sabel and K. Marzullo. Election vs. consensus in asynchronous systems. Technical Report TR95-1488, Cornell University, Feb 1995.

....[3] cannot be solved with 3W ( Gamma) 3. For any Gamma ae Omega Gamma and with reliable channels, we cannot transform any failure detector of P ( Gamma) into some failure detector of P (which implies P ( Gamma) OE P) Hence, for any Gamma ae Omega Gamma problems that need P (e. g election [9], genuine atomic multicast [7] and non blocking atomic commitment [6] cannot be solved with P ( Gamma) The rest of the paper is organized as follows. Section 2 defines the system model. Section 3 defines Gamma accurate failure detectors. Section 4, where we consider eventual reliable ....

....and even with reliable channels, no algorithm can transform any failure detector of P ( Gamma) into some failure detector of 3Q. Hence, we have: P ( Gamma) OE P , Q( Gamma) OE Q, 3P ( Gamma) OE 3P and 3Q( Gamma) OE 3Q. A consequence of P ( Gamma) OE P is that problems requiring P (e. g election [9], genuine atomic multicast [7] and non blocking atomic commitment [6] cannot be solved with P ( Gamma) To proof of P ( Gamma) OE P , Q( Gamma) OE Q, 3P ( Gamma) OE 3P and 3Q( Gamma) OE 3Q, is similar to the proof of the previous section. We introduce a specific failure detector D 0 ( Gamma; ....

L. Sabel and K. Marzullo. Election Vs. Consensus in Asynchronous Systems. Technical Report TR95-1488, Cornell Univ, 1995.

....extending the applicability field of the results of Chandra and Toueg [4] on solving problems in asynchronous systems (with crash failures and reliable channels) augmented with unreliable failure detectors. The applicability of these results to problems other than Consensus has been discussed in [4, 11, 13, 14]. To our knowledge, it is however the first time that (non blocking) atomic commitment problems are discussed in asynchronous systems with unreliable failure detectors. By weakening the NonTriviality condition of atomic commitment, we have defined a problem, called Non Blocking Weak Atomic ....

L. Sabel and K. Marzullo. Election Vs. Consensus in Asynchronous Systems. Technical Report TR95-1488, Cornell Univ, 1995.

.... of different processes do not have to agree (or even eventually agree) Introduced in [12] the abstraction of unreliable failure detectors has been used to solve several important problems such as consensus, atomic broadcast, group membership, non blocking atomic commitment, and leader election [5, 15, 20, 24, 27]. Our goal is to use unreliable failure detectors to achieve quiescence, but before we do so we must address the following important question. Note that any reasonable implementation of a failure detector in a message passing system is itself not quiescent: a process being monitored by a failure ....

....consensus, atomic broadcast, k set agreement, atomic commitment, approximate agreement) to tolerate message losses. It can also be used to extend the results of [6] 4. HB is novel: it is implementable without timeouts, and it does not output lists of suspects as typical failure detectors do [5, 12, 20, 21, 24, 27]. The results of this paper, combined with those in [2] show that lists of suspects is not always the best failure detector output. 5 Reliable communication is a fundamental problem that has been extensively studied, especially in the context of data link protocols (see Chapter 22 of [25] for a ....

L. S. Sabel and K. Marzullo. Election vs. consensus in asynchronous systems. Technical Report 95-1488, Department of Computer Science, Cornell University, Ithaca, New York, Febrary 1995.

....eventual reliable channels, see [1] 3. For any Gamma ae Omega Gamma and with reliable channels, we cannot transform any failure detector of P ( Gamma) into some failure detector of P (which implies P ( Gamma) OE P) Hence, for any Gamma ae Omega Gamma no problem that needs P , e. g election [9], can be solved with P ( Gamma) The rest of the paper is organized as follows. Section 2 defines the system model. Section 3 defines Gamma accurate failure detectors. Section 4, where we consider eventual reliable channels, establishes the above result 1. In Sections 5, 6 and 7, we assume ....

....if j Gammaj 1, W ( Gamma) OE W . 6 About strong accuracy: P ( Gamma) OE P For any subset Gamma ae Omega Gamma and even with reliable channels, no algorithm can transform a failure detector of P ( Gamma) into some failure detector of P . This means that problems requiring P (e. g election [9]) cannot be solved with P ( Gamma) The proof of this result can be found in [7] The idea is very similar to the proof of the previous section. It is based on the notion of frg Gammapattern. We say that a failure pattern F is a frg Gammapattern if only r can crash in F . Using this notion, we ....

L. Sabel and K. Marzullo. Election Vs. Consensus in Asynchronous Systems. Technical Report TR95-1488, Cornell Univ, 1995.

....In this context, consensus would not only be a useful theoretical concept [27, 30] but also a useful service for the clean development of reliable distributed systems. Apart from the agreement problems considered in the paper, one could of course consider other agreement problems like election [24] or terminating reliable broadcast [8] Our framework was designed in the context of asynchronous distributed systems with process crash failures and failure detectors. That is, the framework needs no assumption on process communication delays and process relative speeds. One could apply the ....

L. Sabel and K. Marzullo. Election Vs. Consensus in Asynchronous Systems. Technical Report TR95-1488, Cornell Univ, 1995.

.... have to agree (or even eventually agree) Introduced in [CT96] the abstraction of unreliable failure detectors has been used to solve several important problems such as consensus, atomic broadcast, group membership, non blocking atomic commitment, and leader election [BDM97, DFKM96, Gue95, LH94, SM95] Our goal is to use unreliable failure detectors to achieve quiescence, but before we do so we must address the following important question. Note that any reasonable implementation of a failure detector in a message passing system is itself not quiescent: A process being monitored by a failure ....

....agreement, atomic commitment, approximate agreement) to tolerate message losses. It can also be used to extend the results of [BCBT96] 4. HB is novel: it is implementable without timeouts, and it does not output lists of suspects as typical failure detectors do [BDM97, CT96, Gue95, GLS95, LH94, SM95] The results of this paper, combined with those in [ACT97a] show that lists of suspects is not always the best failure detector output. 5 Reliable communication is a fundamental problem that has been extensively studied, especially in the context of data link protocols (see Chapter 22 of ....

Laura S. Sabel and Keith Marzullo. Election vs. consensus in asynchronous systems. Technical Report 95-1488, Department of Computer Science, Cornell University, Ithaca, New York, Febrary 1995.

....[3] cannot be solved with 3W( Gamma ) 3. For any Gamma ae Omega , and with reliable channels, we cannot transform any failure detector of P( Gamma ) into some failure detector of P (which implies P( Gamma ) OE P) Hence, for any Gamma ae Omega , problems that need P (e. g election [10], genuine atomic multicast [7] and non blocking atomic commitment [6] cannot be solved with P( Gamma ) The rest of the paper is organized as follows. Section 2 defines the system model. Section 3 defines Gamma accurate failure detectors. Section 4, where we consider eventually reliable ....

....even with reliable channels, no algorithm can transform any failure detector of P( Gamma ) into some failure detector of 3Q. Hence, we have: P( Gamma ) OE P, Q( Gamma ) OE Q, 3P( Gamma ) OE 3P, and 3Q( Gamma ) OE 3Q. A consequence of P( Gamma ) OE P is that problems requiring P (e. g election [10], genuine atomic multicast [7] and non blocking atomic commitment [6] cannot be solved with P( Gamma ) The proof of P( Gamma ) OE P, Q( Gamma ) OE Q, 3P( Gamma ) OE 3P and 3Q( Gamma ) OE 3Q, is similar to the proof of the previous section. We introduce a specific failure detector D 0 ....

L. Sabel and K. Marzullo. Election Vs. Consensus in Asynchronous Systems. Technical Report TR95-1488, Cornell Univ, 1995.

....problems, including agreement, mutual exclusion, reliable communication, transaction commit, and clock synchronization. There are quantifiable message complexity results [15] as well as qualitative expressions of impossibility [8, 12] possibility [5, 6] and problems comparative difficulty [5, 10, 24]. Using modal logics of knowledge and time to describe and reason about distributed coordination is well accepted, though unfortunately complex [11, 12, 19, 22, 23, 26] Rarely do these sorts of results find their way to applications programmers; rarer still is the programmer who can spare the ....

L. Sabel and K. Marzullo. Election Vs. Consensus in Asynchronous Systems. Technical Report CS95-411, University of California, San Diego, 1995.

.... only one designated DM needs to be locked, also does not work in an asynchronous system with an unreliable failure detector: the primary copy approach requires to solve the election problem, and the election problem cannot be solved in an asynchronous system with an unreliable failure detector [18]. Intuitively, these difficulties are not surprising since the problems underlying locking and atomic commitment protocols are both agreement problems, i.e. they both require processes to reach a common decision. The so called FLP impossibility result [11] states that in an asynchronous system, ....

L. Sabel and K. Marzullo. Election Vs. Consensus in Asynchronous Systems. Technical Report TR95-1488, Cornell Univ., 1995.

....failure semantics, and 4) processes have no access to hardware clocks. Because in the time free model an observer cannot distinguish between correct, slow or crashed processes, most of the services that are of importance in practice, such as consensus, election or membership, are not implementable [18, 21, 2]. Since [4] we have been using a different model for asynchronous systems, which we have called later [10] the timed asynchronous model, to avoid confusion with the time free model. The timed model assumes that 1) all services are timed: their specification prescribes not only the outputs and ....

....exact specification of the problem, 2) whether the underlying system model allows communication by time, and 3) on the use of progress assumptions. 6.1 Termination vs Conditional Timeliness Conditions There is no commonly agreed upon rigorous specification for the election problem. For example, [21] specifies the election problem for the time free system model as follows: S) at any real time there exists at most one leader, and (TF) infinitely often there exists a leader, i.e. for any real time s there exists a real time t s and a process p so that p becomes leader at t. Problems specified ....

[Article contains additional citation context not shown here]

L. Sabel and K. Marzullo. Election vs. consensus in asynchronous systems. Technical Report TR95-1488, Cornell University, Feb 1995.

....4) processes have no access to hardware clocks. Because in the time free model an observer cannot distinguish between correct, slow or crashed processes, most of the group communication services that are of importance in practice, such as consensus, election or membership, are not implementable [19, 25, 4]. The timed asynchronous system model, introduced without being named in [7] and later named in [16] assumes that 1) all services are timed: their specification prescribes not only the outputs and state transitions that should occur in response to inputs, but also the time intervals within which ....

L. Sabel and K. Marzullo. Election vs. consensus in asynchronous systems. Technical Report TR95-1488, Cornell University, Feb 1995.

No context found.

L. Sabel and K. Marzullo. Election vs. consensus in asynchronous systems. TR 95-1488, Computer Science Department, Cornell University, Ithaca, New York, February 1995.

No context found.

L. Sabel and K. Marzullo. Election vs. consensus in asynchronous systems. TR 95-1488, Computer Science Department, Cornell University, February 1995.

No context found.

L. Sabel and K. Marzullo. Election vs. consensus in asynchronous systems. TR 95-1488, Computer Science Department, Cornell University, Ithaca, New York, February 1995.

No context found.

L. S. Sabel and K. Marzullo. Election vs. consensus in asynchronous systems. Technical report, Cornell University, Ithaca, NY, TR95-1488, 1995.

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