Breaking the Scalability Barrier of Causal Broadcast for Large and Dynamic Systems

Abstract

Many distributed protocols and applications rely on causal broadcast to ensure consistency criteria. However, none of causality tracking state-of-the-art approaches scale in large and dynamic systems. This paper presents a new nonblocking causal broadcast protocol suited for dynamic systems. The proposed protocol outperforms state-of-the-art in size of messages, execution time complexity, and local space complexity. Most importantly, messages piggyback control information the size of which is constant. We prove that for both static and dynamic systems. Consequently, large and dynamic systems can finally afford causal broadcast.

Causal and Δ-causal broadcast in opportunistic networks

“…The new algorithm detailed in [19] also targets large dynamic systems. In this algorithm it is assumed that the processes communicate via an overlay network, whose links are FIFO.…”

“…Since this algorithm does not assume that the underlying system ensures reliable broadcast, the focus is on preventing multiple deliveries of the same message. Like in [19], though, it is assumed that the processes communicate via the FIFO links of an overlay network.…”

Merkle Search Trees: Efficient State-Based CRDTs in Open Networks

“…This algorithm and resulting property is applicable to all CRDTs that are defined by a merge and Merkle Search Trees extend this naturally to maps of CRDTs. This is done without any conditions on the network topology as is required by causal broadcast [6], [26]: it results from the fact of doing a full state merge at every gossip event.…”

“…Causal broadcast is a fundamental primitive of distributed algorithms that is usually directly implemented using send/receive network primitives and vector clocks [5] which are the main alternative to our method and which we show do not scale efficiently. For completeness, let us note that alternative approaches to Causal Broadcast exist for static overlays [26], which have been recently extended to a dynamic setting [6], but the practicality of this latest development has not yet been demonstrated.…”

“…Recent CRDT techniques have focused primarily on leveraging causal broadcast primitives and vector clocks in order to ensure reliable causal delivery of operation deltas [4], [5], meaning that they never need to compare two full states directly. However these approaches show their limit in open networks where many nodes may join and leave, as causal broadcast primitives and vector clocks are both unsuited in this context: causal broadcast has only recently been extended in a scalable fashion to networks with dynamic structures [6] and the practicality of this approach has not yet been demonstrated, whereas vector clocks require metadata that grows linearly with the number of nodes, past and present, that have participated in the network, which is an important scalability barrier even in the case of extensive optimisations [7].…”