Combining static and dynamic analysis to guide PGO for HPC applications: a case study on real-world applications

Modern high performance processor architectures /tackle performance issues by heavily relying on increased vector lengths and advanced memory hierarchies to deliver high performance. This stresses the importance of data access optimization and efficient usage of the underlying hardware. Developers usually trust compilers to automatically address these performance issues, but unfortunately, compilers deploy static performance models and heuristics which, sometimes, remain conservative or even fail in the worst case. Moreover, manual optimization of production HPC codes is not only impractical, but impossible when having to manage multiple architecture dependent transformations. One way to assist compilers is to use Profile Guided Optimization (PGO). It allows the use of feedback data from dynamic profiling using a representative training dataset, for a given target application, enabling the compiler to refine its optimization choices and enhance application performance. But, PGO does not always consider certain metrics and is rarely aggressive enough regarding metric data collection. This bounds the transformation space and limits the compiler's ability to perform further optimizations. An additional option is to provide compilers with user guided assistance in order to enlarge the transformation space (i.e. specialization) and enhance the quality of optimizations.In this paper, we introduce ASSIST, a semi-automatic source-to-source manipulation taking advantage of static and dynamic profiling data produced by performance analysis tools. We demonstrate on real industrial class applications that by combining both static and dynamic analyses and by deploying simple transformations, ASSIST generates similar (and in some cases higher) performance speedups than Intel PGO. Furthermore, combining ASSIST and PGO allows to go a step further, increasing the performance substantially.