Performance and efficiency evaluation and analysis of supercomputers

Performance evaluation is important in supercomputing research and engineering. The widely used Linpack benchmark can only evaluate systems' computing performance. So benchmark suite like HPCC is needed to get an overall performance evaluation. We evaluate four supercomputers using HPCC and get more reasonable answers than using Linpack only. We also introduce two necessary metrics, computing efficiency and energy efficiency, in addition to HPCC to measure systems more comprehensively. Computing efficiency reflects the computation resource utilization rate, and we use quantitative and qualitative methods in combination to evaluate it. Energy efficiency reflects the system's productivity with the unit energy consumption, and MFLOPS/W is the proper metric. Then we show the internal factors affecting system's computing efficiency-the system's organization itself. Efficiencies of systems using different interconnect types are compared to show that interconnect network affects the efficiencies greatly.