Brief Announcement: ArcherGear: Data Race Equivalencing for Expeditious HPC Debugging
Massive parallelization is the driving force for advancements in HPC, but it comes with the cost of increasing code complexity. Programmers need efficient tools to instrument, maintain and debug their codes. An important example is data race detectors: these have the potential to trivialize nondeterministic, confounding bugs. While state of the art race checkers are reasonably efficient in terms of memory and time overhead, they do little to help the programmer analyze the detected races.
In this paper we establish the need for improved race analysis tools, and discuss how aggregating race reports based on various notions of race equivalence can act as the needed mechanism. We also present ArcherGear, a GUI providing most of the described functionality.