Semantics-aware scheduling policies for synchronization determinism

Abstract

A common task for all deterministic multithreading (DMT) systems is to enforce synchronization determinism. However, synchronization determinism has not been the focus of existing DMT research. Instead, most DMT systems focused on how to order data races remained after synchronization determinism is enforced. Consequently, existing scheduling policies for synchronization determinism all have limitations. They may either require performance annotations to achieve good performance or fail to provide schedule stability. In this paper, we argue that synchronization determinism is more fundamental to DMT systems than existing research suggests and propose efficient and effective scheduling policies. Our key insight is that synchronization operations actually encode programmers' intention on how inter-thread communication should be done and can be used as hints while scheduling synchronization operations. Based on this insight, we have built QiThread, a synchronization-determinism system with semantics-aware scheduling policies. Results of a diverse set of 108 programs show that QiThread is able to achieve comparable low overhead as state-of-the-art synchronization-determinism systems without the limitations associated with them.