Performance Implications of Co-scheduling Modern Parallel Applications on NUMA Multi-core Systems

The non-uniform memory access (NUMA) architecture has attracted attention as a state-of-the art multi-core solution for addressing the practical limitations of increasing the number of cores in symmetric multiprocessing systems. In order to utilize this architecture, the scheduling of parallel applications has become an important problem.

In this study, we investigate the performance impact of co-scheduling parallel threads in the recent NUMA platform. From our evaluation, we find that certain applications are found to show significant performance improvements when co-scheduled in the same memory domain. The reason for the improvement is examined, and the use of the memory usage patterns of threads in analyzing the co-scheduling impact is advocated on the basis of the examination result.