Systematic assessment of the overhead of tracing parallel programs

Instant replay is a classical technique developed to help programmers to cope with transient errors occurring in non-deterministic executions of parallel programs. Enough information is recorded during an initial recording phase to be able to force subsequent re-executions to be deterministic with respect to the initial one. If the time overhead of the initial recording is sufficiently low, recording can be used as a normal execution mode of parallel programs. This article describes the method used to assess systematically the overhead of the recording phase of an instant replay tool implemented for a remote procedure call (RPC) based programming model named ATHAPASCAN. Evaluation was done using synthetic programs generated from program models of classical parallel algorithms. The generality of the method comes from the possibility of generating multiple program instances from a single algorithm model. Experimental results confirm the efficiency of the tested execution replay tool