Locality-Aware Task Scheduling and Data Distribution for OpenMP Programs on NUMA Systems and Manycore Processors

Figure 3

Performance impact of data distribution compared to first-touch in programs taken or derived from the Barcelona OpenMP Task Suite (BOTS) [10] and executed on the eight-node Opteron system. Execution time corresponds to the critical path of parallel section. Dispatch stall cycles are aggregated over all program tasks. Most programs improve or maintain performance when data is distributed across NUMA nodes.