A case study of communication trace performance evaluation guided by hierarchical clustering

Today most complex scientific applications requires a large number of calculations to solve a particular problem or set of problems. In order to obtain results in a more rapid fashion, parallel algorithms and applications are developed to run on large-scale supercomputers. However, designing, programming, debugging, and tuning these applications present a challenge. This is in part due to the new level of complexity added by the number of processes that need to explicitly communicate, share data, and synchronize. In addition, it is difficult to track the execution of a program that is being executed simultaneously by multiple processes; this makes it difficult to hand optimize code or to find communication and synchronization bottlenecks. In order to identify potential performance problems, applications are instrumented to collect performance data e.g., hardware performance counter event data and communication traces. This thesis presents a case study of the use of hierarchical clustering to sort through large communication traces and identify processes that potentially have communication bottlenecks or performance problems. (Abstract shortened by UMI.