Tracing Summit 2023 - ThreadMonitor: Low-overhead Data Race Detection using Intel Processor Trace

Data races are among the most difficult bugs to find, due to their non-deterministic nature. With the increasing popularity of multithreaded programming, the need for automated data race detection has significantly increased. Yet, state-of-the-art dynamic data race detectors cannot be used in many real-world testing scenarios, since they cause a significant slowdown and memory overhead. For instance, ThreadSanitizer (TSan) is reported to typically impose a 5×-15× slowdown and 5×-10× memory overhead. This is not tolerable in many industrial testing systems.

This presentation introduces ThreadMonitor (TMon), a low-overhead postmortem data race detector for C/C++ programs. At runtime, it traces the information required for detecting data races (i.e. memory accesses and synchronization events) using Intel Processor Trace (PT), a non-intrusive hardware feature dedicated to tracing software execution. Thereafter, a post-mortem analyzer uses the trace data to emulate the same verification that would be performed by TSan at runtime. TMon has no direct impact on application memory usage and causes a very small slowdown.

The same new technique can be applied to any other architecture or tracing facility, provided that the required information can still be traced. Also, a similar approach may be adapted to design post-mortem tools that emulate other runtime verification tools, such as AddressSanitizer (ASan).