Auto-tuning of OpenMP Applications on the IBM Blue Gene/Q

Modern high performance computing architectures are based on multi-core and multi-threaded computing nodes. The mixed MPI and OpenMP programming is currently a reference model for obtaining high scalability on large computing systems. In such a model, MPI processes contain many OpenMP parallel regions. Scalability and performance of those parallel regions may di er between various computing systems and between each run of the code. The control of the number of threads used by di erent OpenMP regions, by users of the HPC systems, is very often limited to setting a single environment variable - OMP NUM THREADS. In this work we present a tool called SOMPARlib which is based on OpenMP Monitoring Interface (POMP) and is capable of controlling the execution of various OpenMP parallel regions introduced in computational codes during run time. The tool is particularly useful in the case of architec- tures that introduce the multithreading mechanisms like Simultaneous multithreading (SMT) or Hyper-Threading (HT)