Partially ordered distributed computations on asynchronous point-to-point networks

TL;DR

This paper introduces partially ordered execution models for asynchronous distributed algorithms, aiming to identify more efficient execution patterns in event- and pulse-driven computations, with specific algorithms for backtrack search and linear systems.

Contribution

It defines partially ordered execution frameworks for asynchronous distributed algorithms and provides algorithms ensuring these restricted executions, improving efficiency for certain applications.

Findings

Partially ordered executions can be more efficient in specific distributed tasks.

Algorithms are proposed to ensure these restricted execution patterns.

Applications include backtrack search and iterative linear algebra algorithms.

Abstract

Asynchronous executions of a distributed algorithm differ from each other due to the nondeterminism in the order in which the messages exchanged are handled. In many situations of interest, the asynchronous executions induced by restricting nondeterminism are more efficient, in an application-specific sense, than the others. In this work, we define partially ordered executions of a distributed algorithm as the executions satisfying some restricted orders of their actions in two different frameworks, those of the so-called event- and pulse-driven computations. The aim of these restrictions is to characterize asynchronous executions that are likely to be more efficient for some important classes of applications. Also, an asynchronous algorithm that ensures the occurrence of partially ordered executions is given for each case. Two of the applications that we believe may benefit from the restricted nondeterminism are backtrack search, in the event-driven case, and iterative algorithms for systems of linear equations, in the pulse-driven case.