Causing Communication Closure: Safe Program Composition with Reliable Non-FIFO Channels

TL;DR

This paper introduces a semantic framework for safe composition of distributed programs over reliable non-FIFO channels, emphasizing the importance of sealing to ensure communication-closed execution without message overtaking.

Contribution

It defines the concept of sealing for safe program composition in non-FIFO channels and provides algorithms for testing and constructing seals, connecting Lamport causality with composition safety.

Findings

Sealing ensures programs run in isolation as if in a closed environment.

Algorithms are provided for testing and constructing seals efficiently.

Every sealable program with interference on O(n^2) channels can be sealed with O(n) messages.

Abstract

A semantic framework for analyzing safe composition of distributed programs is presented. Its applicability is illustrated by a study of program composition when communication is reliable but not necessarily FIFO\@. In this model, special care must be taken to ensure that messages do not accidentally overtake one another in the composed program. We show that barriers do not exist in this model. Indeed, no program that sends or receives messages can automatically be composed with arbitrary programs without jeopardizing their intended behavior. Safety of composition becomes context-sensitive and new tools are needed for ensuring it. A notion of \emph{sealing} is defined, where if a program $P$ is immediately followed by a program $Q$ that seals $P$ then $P$ will be communication-closed--it will execute as if it runs in isolation. The investigation of sealing in this model reveals a novel connection between Lamport causality and safe composition. A characterization of sealable programs is given, as well as efficient algorithms for testing if $Q$ seals $P$ and for constructing a seal for a significant class of programs. It is shown that every sealable program that is open to interference on $O(n^2)$ channels can be sealed using O(n) messages.