Performance prediction of parallel systems based on workload similarity

Performance prediction of workloads on parallel systems use a priori information to estimate performance when the input data size, or the machine parameters, change. This work fills an important gap. Given an application that has never been implemented on a target machine, we propose a methodology to predict the performance of such an application on that machine. This allows application developers to make intelligent choices before committing to a specific machine, directly without having their own benchmarking activity. This is accomplished by representing the workloads using the parallel instruction centroid, which is a metric that embodies parallelism, critical path length, and instruction 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 communication requirements. In addition to filling the previously described gap, it will be shown that this method provides higher prediction accuracy in the majority of the cases, and accounts for dynamic code behaviors.