Modeling cost/performance of a parallel computer simulator

Reviewer: Guenter Haring

The cost/performance ratio of simulating a hypothetical target parallel computer at system level using a commercial host parallel computer is examined. The authors give a comprehensive description of the performance model used in previous papers. They develop an analytic performance model of the Wisconsin Wind Tunnel (WWT) to accurately predict the simulation speedup, which is the ratio of the time for simulating N target nodes on p host nodes and on one host node. The WWT simulates cache-coherent shared-memory machines on a message-passing CM-5 using direct execution, which exploits the commonality between the instruction sets of the simulated and the host computer. In WWT, the interprocess interactions are managed by dividing program execution into lock-step quanta. The model, described in the main part of the paper, predicts the running time of a quantum. To accurately model the WWT's performance, the running time is broken down on the critical path into four components: the sum of the processing times, the direct overheads, the indirect overheads, and the quantum synchronization time. The variability in event processing time is modeled using a slightly modified version of Kruskal and Weiss's model for fork-join parallel programs based on standard analysis-of-variance techniques. WWT uses separate address space for each target node, causing direct context switch overhead, the frequency of which is modeled by the maximum of binomial random variables. To predict the interference of multiple targets in the host cache and TLB (indirect overhead), a generalization of Thie`baut and Stone's footprint model is used, by allowing sharing among address spaces of multiple (more than two) processes. The performance model extracts parameters from a fully parallel simulation of a specific target system and uses them to predict the performance of the same simulation running on a different number of host nodes. The model was validated using five different benchmarks. A more detailed explanation of the observed results would have improved this part of the paper. The part on modeling cost/performance starts by describing the assumptions about how the target and the simulating system scale, and their consequences for the model. Then the authors introduce simple cost models for uniprocessors, small-scale bus-based shared-memory multiprocessors, and large-scale parallel supercomputers. These models allow them to perform a detailed cost/performance analysis by varying the number of host processors and the number of target nodes per host node. With this kind of analysis, the optimum number K=N/p of target nodes simulated per host node can be found. K is essentially independent of p . Furthermore, the authors show under which conditions the parallel simulation is not only faster but more cost-effective than sequential simulation. In the examples used, the authors give the cost-effectiveness for target systems that have more than 32 nodes.