Performance Metrics and Measurement Techniques of Collective Communication Services

The performance of collective communication is critical to the overall system performance. In general, the performance of the collective communication is dependent not only on the underlying hardware, but also its implementation. To evaluate the performance of collective communication accurately, identifying the most representative metrics and using correct measurement techniques are two important issues that have to be considered. This paper focuses on the measurement techniques for collective communication services. The proposed techniques can provide an accurate evaluation of the completion time without requiring a global clock and without having to know the detailed implementations of collective communication services. Experimental results conducted on the IBM/SP at Argonne National Laboratory are presented.