Practical Byzantine Group Communication

Abstract

This paper presents an adaptation of a group communication system called JazzEnsemble to tolerate Byzantine failures. The work described here emphasizes scalability and good performance in the normal case, i.e., when there are no failures, while providing strong semantics to the application. The paper presents the main concepts and protocols that enable the Byzantine tolerant version of JazzEnsemble to obtain these goals. In particular, this includes fuzzy mute and fuzzy verbose failure detectors, an efficient Byzantine vector consensus protocol, and a novel Byzantine uniform broadcast protocol, as well as modifications at each layer of the system to overcome potential Byzantine attacks. Additionally, high-level protocols only rely on the oral messages model, and thus messages need to be signed only once at a low level of the system. Finally, the paper presents an extensive performance evaluation, which demonstrates the system's scalability and efficiency. This is also used to analyze the sources of performance degradation associated with various aspects of overcoming Byzantine failures.

On the Reduction of Atomic Broadcast to Consensus with Byzantine Faults

“…Although there is a huge amount of literature on the implementation of atomic broadcast under Byzantine faults, and some of them even use some sort of agreement as a building block, most of these papers do not-strictly speaking-reduce atomic broadcast to consensus. 1 In fact, we are aware only of three reductions of atomic broadcast to consensus with Byzantine faults [2], [3], [4], and in our opinion this work does not fully clarify the relation of consensus and atomic broadcast.…”

Prime: Byzantine Replication under Attack

“…Both protocols disseminate requests to 2f +1 servers before a coordinator assigns the global order. Drabkin et al [26] observe the difficulty of setting timeouts in the context of group communication in malicious settings. Prime's Reconciliation sub-protocol uses erasure codes for efficient data dissemination.…”

The Fail-Heterogeneous Architectural Model

“…As many other fail-crash tolerant protocols, group membership has been adapted to operate in Byzantine environments (as in [10]). However, in order to prevent malicious nodes from triggering isolations of correct nodes it is necessary that a process is accused by at least f + 1 other processes, where f is the number of Byzantine nodes, before a view change is initiated.…”