Automatic construction of coordinated performance skeletons

Performance prediction is particularly challenging for dynamic and unpredictable environments that cannot be modeled well, such as execution with sharing of CPU and bandwidth resources. Our approach to performance estimation in such scenarios is based on actual execution of short running customized performance skeletons for target applications. This work focuses on automatic construction of performance skeletons for parallel MPI programs. Logicalization of a family of traces to a single trace is presented as a key technique for skeleton construction. Compression of communication traces is achieved by identifying the loop structure from traces. Results are presented that demonstrate that logicalization and compression are accurate and efficient. Automatically constructed performance skeletons were able to effectively predict application performance in a variety of scenarios involving resource sharing and changes in the execution environment.