An efficient record-replay mechanism for shared memory programs

One of the main difficulties in debugging parallel programs is that parallel programs execute non-deterministically. As a consequence, cyclic debugging techniques for error isolation are not guaranteed to work. For that reason, different record-replay mechanisms have been proposed. Ordering based mechanisms log the relative order in which process interactions take place in order to obtain reproducibility. In this paper, existing ordering based approaches for shared memory programs are compared, and a novel approach, situated somewhere in between the existing approaches, is presented. To obtain repeatability, we record the relative ordering of the processor-level synchronization events. The proposed mechanism can also be used in support of performance debugging