H.M. Jafri (USA)
Performance analysis, root-cause analysis, parallel com puting.
Parallel applications are notorious for their intractability to performance debugging. Automatic performance analysis techniques, such as those used by Kojak and KappaPI, are promising in alleviating the difficulty of discovering per formance inefficiencies in parallel applications. However, as we show in this paper, the results produced by these tool can be potentially misleading and sometimes, outright in correct. The reason is that the overhead due to performance inefficiencies originating at a certain point in the program can causally propagate and manifest itself at other points. Current techniques perform a flat analysis, i.e., they do not account for causal propagation. In this paper, we present a method of causal analysis that current analysis techniques can be retrofitted with to account for causal propagation of overhead to arrive at a more accurate description of per formance bottlenecks. We also show various advantages rendered by this technique to improving the effectiveness of automatic performance analysis. In this paper, we only tackle overhead related to communication operations in MPI parallel application. In general, however, our tech nique can be used for non-communication related overhead for any parallel programming paradigm.
Important Links:
Go Back
