A Hardware Approach to Detect, Expose and Tolerate High Level Data Races

Concurrent programs are more complex and error prone than their sequential peers, and are much harder to debug as well. High level data races (HLDR) are one of the concurrency bugs most difficult to debug. They are a class of concurrency errors that are not commonly addressed by the testing and debugging techniques and tools. HLDR result from the misdefinition of the scope of an atomic block, which should be unique but was wrongly split into two or more independent atomic blocks. Interleavings involving these misdefined atomic blocks may violate the program correctness invariants and cause the concurrent program to fail. In this work we propose a hardware module to detect, expose and tolerate HLDR in concurrent programs, with applications in both the software testing and debugging and the software deployment phases. In the detecting mode, our proposal detects HLDR with few false positives and without the overhead and intrusion of other dynamic software approaches. In the exposing mode, it "stimulates" the program to expose existing latent HLDR and trigger hidden HLDRs. Finally, in the tolerating mode, it may act as a software healing technique by inhibiting certain buggy interleavings. The results shows a reasonable performance overhead and few false positives in all modes.