A compiler approach to performance prediction using empirical-based modeling

Performance understanding and prediction are extremely important goals for guiding the application of program optimizations or in helping programmers focus their efforts when tuning their applications. In this paper we survey current approaches in performance understanding and modeling for high-performance scientific applications. We also describe a performance modeling and prediction approach that relies on the synergistic collaboration of compiler analysis, compiler-generated instrumentation (to observe relevant run-time input values) and multimodel performance modeling. A compiler analyzes the source code to derive a discrete set of parameterizable performance models. The models use run-time data to define the values of their parameters. This approach, we believe, will allow for higher performance modeling accuracy and more importantly to more precise identification of what the causes of performance problems are.