Evaluation of Early Systems -- recent results
Evaluation of Early Systems -- recent results
Communication Protocol Comparison Results
As part of the Evaluation of Early Systems project and the CHAMMP climate modeling program, benchmarks drawn from the Parallel Community Climate Model PCCM2 are being used to evaluate parallel machines of interest to the U.S. High Performance Computing and Communications (HPCC) Program. One focus of this research is on how best to use each parallel platform.
Over the past two years, studies have been performed to identify the optimal parallel algorithms and algorithm implementations to use for the Parallel Spectral Transforms Shallow Water Model (PSTSWM). Many of the results on optimal implementations and communication protocols from these studies are independent of the application, and represent general advice for code developers on how to implement parallel algorithms.
The following tables, from a comparison completed in April 1994, describe the variation in execution time for the entire application code when varying the communication protocols used in different parallel FFT algorithms. The table describes the execution time of the slowest protocol (MAX) and of the slowest protocol in the fastest 25% (Q1), both relative to the time of the fastest algorithm, indicating how important tuning is and how sensitive performance is to getting the optimal protocol. Results for two different problem sizes and for the nCUBE 2 (N), iPSC/860 (I), Paragon XP/S using OSF (P), and Paragon XP/S using SUNMOS (S) are reported. It is important to note in these results that the optimal protocol varies (dramatically) between the different machines.
Worley and Foster, "Parallel Spectral Transform Shallow Water Model: A runtime-tunable parallel benchmark code".
Data for a similar study have also been collected for the Intel Paragon XP/S with current versions of the hardware and operating systems, for the Cray Research Inc. T3D, and for the IBM SP-2, and is currently being analyzed.
Evaluation of Early Systems Project