Piecewise holistic autotuning of parallel programs with CERE

摘要

Current architecture complexity requires fine tuning of compiler and runtime parameters tornachieve best performance.Autotuning substantially improves default parameters inmany scenarios,rnbut it is a costly process requiring long iterative evaluations.We propose an automatic piecewisernautotuner based on CERE (Codelet Extractor and Replayer). CERE decomposes applicationsrninto small pieces called codelets: Each codeletmaps to a loop or to anOpenMPparallel region andrncan be replayed as a standalone program.Codelet autotuning achieves better speedups at a lowerrntuning cost. By grouping codelet invocations with the same performance behavior, CERE reducesrnthe number of loops or OpenMP regions to be evaluated. Moreover, unlike whole-program tuning,rnCERE customizes the set of best parameters for each specific OpenMP region or loop. Werndemonstrate the CERE tuning of compiler optimizations, number of threads, thread affinity, andrnscheduling policy on both nonuniformmemoryaccess and heterogeneous architectures.Over thernNAS benchmarks, we achieve an average speedup of 1.08× after tuning. Tuning a codelet is 13×rncheaper than whole-program evaluation and predicts the tuning impact with a 94.7% accuracy.rnSimilarly, exploring thread configurations and scheduling policies for a Black-Scholes solver on anrnheterogeneous big.LITTLE architecture is over 40× faster using CERE.