Optimal Prediction of Synchronization-Preserving Races

TL;DR

This paper introduces a new class of race prediction called sync-preserving races, providing a sound, complete, and nearly optimal algorithm for detecting them in concurrent programs, and establishing their computational complexity boundaries.

Contribution

It defines sync-preserving races, develops an optimal algorithm for predicting them, and proves the problem's computational complexity limits, advancing race detection techniques.

Findings

Algorithm runs in near-linear time and space for moderate parameters.

Predicting races with a single sync reversal is NP-complete.

The problem has a matching lower bound, indicating optimality.

Abstract

Concurrent programs are notoriously hard to write correctly, as scheduling nondeterminism introduces subtle errors that are both hard to detect and to reproduce. The most common concurrency errors are (data) races, which occur when memory-conflicting actions are executed concurrently. Consequently, considerable effort has been made towards developing efficient techniques for race detection. The most common approach is dynamic race prediction: given an observed, race-free trace $\sigma$ of a concurrent program, the task is to decide whether events of $\sigma$ can be correctly reordered to a trace $\sigma^*$ that witnesses a race hidden in $\sigma$. In this work we introduce the notion of sync(hronization)-preserving races. A sync-preserving race occurs in $\sigma$ when there is a witness $\sigma^*$ in which synchronization operations (e.g., acquisition and release of locks) appear in the same order as in $\sigma$. This is a broad definition that strictly subsumes the famous notion of happens-before races. Our main results are as follows. First, we develop a sound and complete algorithm for predicting sync-preserving races. For moderate values of parameters like the number of threads, the algorithm runs in $\widetilde{O}(\mathcal{N})$ time and space, where $\mathcal{N}$ is the length of the trace $\sigma$. Second, we show that the problem has a $\Omega(\mathcal{N}/\log^2 \mathcal{N})$ space lower bound, and thus our algorithm is essentially time and space optimal. Third, we show that predicting races with even just a single reversal of two sync operations is $\operatorname{NP}$-complete and even $\operatorname{W}[1]$-hard when parameterized by the number of threads. Thus, sync-preservation characterizes exactly the tractability boundary of race prediction, and our algorithm is nearly optimal for the tractable side.