Performance evaluation of CP-PACS on CG benchmark
K. Itakura, Hiroshi Nakamura, Taisuke Boku, K. Nakazawa · 2002
We evaluate NAS Parallel Benchmarks ver.1 Kernel CG on massively parallel processor CP-PACS, and analyze the results. CP-PACS' CPU has a special register which is auto-incremented by clock cycle, and we can measure the time spent on any function routine with very high accuracy. As a result of the performance analysis, especially of the data transfer time, our desk-top estimation fits the measured results almost perfectly. From this analysis, we can show the program bottlenecks when executing with a large number of PUs.