PERFORMANCE PREDICTION OF PARALLEL SYSTEMS BASED ON WORKLOAD SIMILARITY

Performance prediction of workloads on parallel systems use a priori i nformation to estimate performance when the input data size, or the ma chine parameters, change. This work fills an important gap. Given an a pplication that has never been implemented on a target machine, we pro pose a methodology to predict the performance of such an application o n that machine. This allows application developers to make intelligent choices before committing to a specific machine, directly without hav ing their own benchmarking activity. This is accomplished by represent ing the workloads using the parallel instruction centroid, which is a metric that embodies parallelism, critical path length, and instructio n mixes properties. The difference between these centroids is measured as a representation of similarity. The most similar workload to ours is used for prediction, after compensating for the difference in commu nication requirements. In addition to filling the previously described gap, it will be shown that this method provides higher prediction acc uracy in the majority of the cases, and accounts for dynamic code beha viors.