Optimal Asynchronous Garbage Collection for RDT Checkpointing Protocols

Schmidt, Roman; Garcia, Islene C.; Pedone, Fernando; Buzato, Luiz Eduardo

doi:10.1109/icdcs.2005.58

Cited by 4 publications

(4 citation statements)

References 23 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…This is particularly important since it has been shown that n is also the upper bound on the number of nonobsolete checkpoints per process [1]. In [2] we present an algorithm based on this condition that eliminates checkpoints on-the-fly during execution. It runs locally to each process and is based only on the timestamps already propagated by the checkpointing protocol, increasing neither the amount of control information piggybacked nor the execution complexity of the checkpointing middleware.…”

mentioning

confidence: 99%

“…The paper shows that it is possible to do asynchronous garbage collection when RDT holds [2]. We introduce a necessary and sufficient condition to make a checkpoint obsolete in RDT scenarios and derive from that a sufficient condition to identify obsolete checkpoints based only on causal knowledge.…”

mentioning

confidence: 99%

“…If no such checkpoint exists, we assume lk i(j) = −1. Theorem 1 presents a condition that allows asynchronous garbage collection in RDT scenarios [2]. Although not necessary, this condition has an important practical implication: based on it a process retains a maximum of n stored checkpoints, where n is the number of processes.…”

mentioning

confidence: 99%

See 2 more Smart Citations

Brief announcement

Schmidt

Garcia

Pedone

et al. 2004

Proceedings of the Twenty-Third Annual ACM Symposium on Principles of Distributed Computing

View full text Add to dashboard Cite

Checkpointing is a well-known technique to build faulttolerant distributed applications based on rollback-recovery. Briefly, every process periodically saves the application's local state as a checkpoint and when a failure occurs, the distributed computation restarts from its most recent consistent global checkpoint, or recovery line. If checkpoints are taken autonomously by processes (called basic checkpoints) and no coordination exists, a failure could force the application to roll back to a very initial state. This happens because checkpoint dependencies can be non-causal and make a checkpoint depend on itself, not being part of any consistent global checkpoint. Communication-induced checkpointing protocols avoid this problem by piggybacking control information in the application messages and having processes take extra forced checkpoints. Some communicationinduced checkpointing protocols ensure that all checkpoint dependencies are causal and, therefore, can be tracked online during the application execution. This property, called rollback-dependency trackability (RDT), provides important guaranties for the recovery system and has been shown to be helpful in many contexts [3].The price of autonomy in communication-induced checkpointing protocols is storage space. The absence of explicit coordination makes it difficult to identify obsolete checkpoints, that is, those not necessary for future recoveries. Although garbage collection incurs overhead, being an important pragmatic issue in rollback-recovery, it has received little attention in the literature [1]. Existent algorithms eliminate all or a subset of the obsolete checkpoints relying either on time assumptions or reliable control message exchanges. Ideally, garbage collection should be as little intrusive as Copyright is held by the author/owner. PODC'04, July 25-28, 2004, St. Johns, Newfoundland, Canada. ACM 1-58113-802-4/04/0007. possible, not introducing any overhead in the normal computation or blocking the execution of processes. We capture this intuition with the notion of asynchronous garbage collection, which relies only on information piggybacked in the application messages.The paper shows that it is possible to do asynchronous garbage collection when RDT holds [2]. We introduce a necessary and sufficient condition to make a checkpoint obsolete in RDT scenarios and derive from that a sufficient condition to identify obsolete checkpoints based only on causal knowledge. Let c γ i represent the γ-th checkpoint taken by process pi, where γ (≥ 0) is its index, and lk i(j) denote the index of the last checkpoint of process pj known by process pi, that is, the last checkpoint of pj that causally precedes the current state of pi. If no such checkpoint exists, we assume lk i(j) = −1. Theorem 1 presents a condition that allows asynchronous garbage collection in RDT scenarios [2]. Theorem 1. A checkpoint c γ i is obsolete if there is no process p f such that lk i(f ) ≥ 0 and c γ i is the last checkpoint of pi not causally preceded by c lk i (f ) f .Although ...

show abstract

mentioning

confidence: 99%

mentioning

confidence: 99%

See 1 more Smart Citation

Brief announcement

Schmidt

Garcia

Pedone

et al. 2004

Proceedings of the Twenty-Third Annual ACM Symposium on Principles of Distributed Computing

View full text Add to dashboard Cite

show abstract

“…Em particular, a coleta de lixo ingênua ou a ótima em padrões de checkpoints e mensagens que satisfazem a propriedade RDT podem ser simplificadas. Neste caso, os processos devem manter, no máximo, n checkpoints e a linha de recuperação pode ser calculada apenas com informações locais sem a necessidade de troca de mensagens ou bloqueio dos processos durante sua execução [37].…”

Section: Coleta De Lixounclassified

Uma ponte entre as abordagens sincrona e quase-sincrona para checkpointing

Sakata¹

View full text Add to dashboard Cite

Protocolos de checkpointing são responsáveis pelo armazenamento de estados dos processos de um sistema distribuído em memória estável para tolerar falhas. Os protocolos síncronos minimais induzem apenas um número minimal de processos a salvarem checkpoints durante uma execução do protocolo bloqueando os processos envolvidos. Uma versão não-bloqueante desta abordagem garante a minimalidade no número de checkpoints salvos em memória estável com o uso de checkpoints mutáveis, checkpoints que podem ser salvos em memória não-estável. Porém, a complexidade deste protocolo e o fato de ele tolerar apenas a presença de uma execução de checkpointing a cada instante nos motivou a procurar soluções para estes problemas na teoria desenvolvida para os protocolos quase-síncronos.A nova abordagem nos permitiu fazer uma revisão de alguns protocolos síncronos bloqueantes existentes na literatura que até então eram considerados minimais. Nesta mesma linha, obtivemos novos resultados na análise de minimalidade dos protocolos síncronos não-bloqueantes, ao considerarmos a aplicação como um todo e também a existência de execuções concorrentes de checkpointing.Ao estabelecermos esta ponte entre as abordagens para checkpointing, conseguimos desenvolver dois novos protocolos síncronos não-bloqueantes. Ambos fazem uso de checkpoints mutáveis, permitem execuções concorrentes de checkpointing e possuem um mecanismo simples de coleta de lixo. No entanto, o fato de cada um dos protocolos derivar de classes diferentes de protocolos quase-síncronos leva a comportamentos distintos, como evidenciado por resultados de simulação. vii

show abstract