Performance Modelling of Message-Passing Parallel Programs

Abstract

Parallel computing is essential for solving very large scientific and engineering problems. An effective parallel computing solution requires an appropriate parallel machine and a well-optimised parallel program, both of which can be selected via performance modelling. This dissertation describes a new performance modelling system, called the Performance Evaluating Virtual Parallel Machine (PEVPM). Unlike previous techniques, the PEVPM system is relatively easy to use, inexpensive to apply and extremely accurate. It uses a novel bottom-up approach, where submodels of individual computation and communication events are dynamically constructed from data-dependencies, current contention levels and the performance distributions of low-level operations, which define performance variability in the face of contention. During model evaluation, the performance distribution attached to each submodel is sampled using Monte Carlo techniques, thus simulating the effects of contention. This allows the PEVPM to accurately simulate a program's execution structure, even if it is non-deterministic, and thus to predict its performance.

Obtaining these performance distributions required the development of a new benchmarking tool, called MPIBench. Unlike previous tools, which simply measure average message-passing time over a large number of repeated message transfers, MPIBench uses a highly accurate and globally synchronised clock to measure the performance of individual communication operations. MPIBench was used to benchmark three parallel computers, which encompassed a wide range of network performance capabilities, namely those provided by Fast Ethernet, Myrinet and QsNet. Network contention, a problem ignored by most research in this area, was found to cause extensive performance variation during message-passing operations. For point-to-point communication, this variation was best described by Pearson 5 distributions. Collective communication operations were able to be modelled using their constituent point-to-point operations. In cases of severe contention, extreme outliers were common in the observed performance distributions, which were shown to be the result of lost messages and their subsequent retransmit timeouts.

The highly accurate benchmark results provided by MPIBench were coupled with the PEVPM models of a range of parallel programs, and simulated by the PEVPM. These case studies proved that, unlike previous modelling approaches, the PEVPM technique successfully unites generality, flexibility, cost-effectiveness and accuracy in one performance modelling system for parallel programs. This makes it avaluable tool for the development of parallel computing solutions.