Trace size vs parallelism in trace-and-replay debugging of shared-memory programs

Abstract

Execution replay is a debugging strategy where a program is run repeatedly on an input that manifests bugs. Replaying nondeterministic parallel programs requires special tools; otherwise, successive runs (on the same input) can differ, making bugs impossible to track. These tools must trace an execution so it can be replayed. We present improvements over our past work on an adaptive tracing strategy for shared-memory programs. Our past approach makes run-time tracing decisions by detecting and tracing exactly the non-transitive dynamic data dependences among the execution's shared data. Tracing the non-transitive dependences provides sufficient information for a replay. In this paper we show that tracing exactly these dependences is not necessary. Instead, we present two algorithms that introduce and trace artificial dependences among some events that are actually independent If no data dependence exists between two memory references during execution, we are free to artificially force them to execute in a specific order during replay. Artificial dependences reduce trace size, but introduce additional event orderings that have the potential of reducing the replay's parallelism. We present one algorithm that always adds dependences guaranteed not to be on the critical path (which do not slow replay). Another algorithm adds as many dependences as possible, slowing replay but reducing trace size further. Experiments show that we can improve the already high trace reduction of our past technique by up to two more orders of magnitude, without slowing replay. Our new techniques usually trace only 0.00025–0.2% of the shared-memory references, a 3–6 order of magnitude reduction over past approaches that trace every access.

This research was partly supported by ONR Contract N00014-91-J-4052 (ARPA Order 8225) and NSF grant CCR-9309311.