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

Figure 9

Performance of data distribution combined with work-stealing and locality-aware scheduling on TILEPro64. Execution time is normalized to performance of work-stealing scheduling with fine distribution for each benchmark. Inputs to Map: 63 integer vectors, 32 kB each; Reduction: 700 kB integer array and depth = 6; Vecmul: 128 integer vectors, 28 kB each; SparseLU: 1152 × 1152 integer matrix, block size = 36, and intensity heuristic enabled for tasks executing the bmod function. Locality-aware scheduling, in combination with heuristic-guided data distribution, improves or maintains performance compared to work-stealing.