Solving Atomic Multicast When Groups Crash

Abstract

In this paper, we study the atomic multicast problem, a fundamental abstraction for building faulttolerant systems. In the atomic multicast problem, the system is divided into non-empty and disjoint groups of processes. Multicast messages may be addressed to any subset of groups, each message possibly being multicast to a different subset. Several papers previously studied this problem either in local area networks [3,9,20] or wide area networks [13,21]. However, none of them considered atomic multicast when groups may crash. We present two atomic multicast algorithms that tolerate the crash of groups. The first algorithm tolerates an arbitrary number of failures, is genuine (i.e., to deliver a message m, only addressees of m are involved in the protocol), and uses the perfect failures detector P. We show that among realistic failure detectors, i.e., those that do not predict the future, P is necessary to solve genuine atomic multicast if we do not bound the number of processes that may fail. Thus, P is the weakest realistic failure detector for solving genuine atomic multicast when an arbitrary number of processes may crash. Our second algorithm is non-genuine and less resilient to process failures than the first algorithm but has several advantages: (i) it requires perfect failure detection within groups only, and not across the system, (ii) as we show in the paper it can be modified to rely on unreliable failure detection at the cost of a weaker liveness guarantee, and (iii) it is fast, messages addressed to multiple groups may be delivered within two inter-group message delays only.

P-Store: Genuine Partial Replication in Wide Area Networks

“…We assume that our system is equipped with an atomic multicast service that allows messages to be disseminated to any subset of groups in Γ [11], [12]. For every message m, m.dst denotes the groups to which m is multicast.…”

“…We assumed that groups were correct and used an atomic multicast service optimized for this assumption. To provide higher degrees of resilience, it would have sufficed to replace the atomic multicast service with an implementation that tolerates group crashes [12] and replicate data items in multiple groups. Note, however, that this would come at a cost.…”

Fast, flexible, and highly resilient genuine fifo and causal multicast algorithms

“…Fifo and causal order help the programming of distributed applications in various domains such as global snapshot construction [2] and fair resource allocation [10]. Causal multicast may also serve as a building block to implement atomic multicast [16].…”

Genuine versus Non-Genuine Atomic Multicast Protocols for Wide Area Networks: An Empirical Study

“…We identify two existing inter-group-latency optimal protocols: the disastervulnerable genuine algorithm in [16], which assumes that each group contains one correct process, and the disastertolerant algorithm in [17]. We also introduce a non-genuine algorithm that is an optimization of [17] for the case of correct groups, i.e., groups that contain at least one correct process. The resulting algorithm may deliver global messages in a single-group delay (i.e., it is inter-group-latency optimal) and may deliver local messages with no inter-group communication.…”

“…We then propose techniques to reduce this effect in [16,17] as well as the non-genuine algorithm proposed in this paper. Although simple, these techniques decrease the delivery latency of local messages by as much as two orders of magnitude.…”