CORD: A Collaborative Framework for Distributed Data Race Detection

Modern concurrent software is riddled with data races and these races constitute the source of many problems. Data races are hard to detect accurately before software is shipped and, once they cause failures in production, developers find it challenging to reproduce and debug them. 

Ideally, all data races should be known before software ships. Static data race detectors are fast, have few false negatives, but unfortunately have many false positives. Conversely, dynamic data race detectors do not have false positives, but have many false negatives and incur high runtime overhead. There is no silver bullet and, as a result, modern software still ships with numerous data races.

We present CoRD, a collaborative distributed testing framework that aims to combine the best of the two approaches: CoRD first statically detects races and then dynamically validates them via crowdsourced executions of the program. Our initial results show that CoRD is more effective than static or dynamic detectors alone, and it introduces negligible runtime overhead.