Dynamic Tuning of OpenMP Memory Bound Applications in Multisocket Systems using MATE

The performance of OpenMP applications executed in multisocket multicore processors can be limited by the memory interface. In a multisocket environment, each multicore processor can present a performance degradation in memory-bound parallel regions when sharing the same Last Level Cache (LLC). In this case the use of all available resources is not always the best choice in terms of execution time and/or efficiency. The best configuration for an application depends on the system's architecture, the input data, and the data evolution; hence, it can vary from execution to execution or even during the same execution. This means that, in order to find a configuration that makes an efficient use of the available resources, an adequate methodology and tools are needed. In this work we present the integration of a performance model for OpenMP memory bound applications in a dynamic performance tuning tool called MATE. For achieving this integration, MATE was extended to support measurement of hardware counters, and the performance model was adapted for determining the best number of threads for an application and for being implemented in MATE as a Tunlet.

The developed Tunlet has been evaluated using different multi-socket architectures and memory bound application benchmarks, showing that the proposed approach can be efficient.