Performance of ALE3D on the ASCI machines
Evi Dube, Scott Futral, J. Robert Neely, Taylor Pierce · University of North Texas Digital Library (University of North Texas) · 1999
With the anticipated delivery of the ASCI Blue Pacific SST machine approaching, the scaling and performance on large numbers of processors for B Division applications codes have become a matter of consider able interest. Besides the practical impact on users (achievable problem size and runtime), the application codes' performance are the ultimate measure of th esuccess of the ASCI machines.The ALE3D code was used to evaluate the performance of the current Blue PacificTechnical Refresh hardware and software in various modes of running. We will present results and analysis of the performance behavior from this study, along with results from other ASCI machines.While gathering statics from user problem runs is easy to do, it is difficult to analyze the variation in performance from problem to problem, or to adjust the problem size consistently for scaling studies. Trivial problems can be used, but may not well reflect the actual performance users can expect. For this study, a series of problems were used that reflect the characteristics of real user problems, but allow for uniform, constant work per processor scaling of problem size and well understood communication characteristics between processors. Additional results for fixed-size problems are presented. Runs were done using different message passing modes (User Space, I.P. protocol), processors per node, and environment settings to investigate the IBM SP2 performance characteristics.