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

Figure 7

Performance of data distribution combined with work-stealing and locality-aware scheduling on eight-node Opteron system. Execution time is normalized to performance of work-stealing with memory page interleaving using numactl for each benchmark. Inputs to Map: 48 floating-point vectors, 1 MB each; Jacobi: 16384 × 16384 floating-point matrix and block size = 512; Matmul: 4096 × 4096 floating-point matrix and block size = 128; SparseLU: 8192 × 8192 floating-point matrix and block size = 256; Reduction: 256 MB floating-point array and depth = 10. Combination of numactl page-wise interleaving and locality-aware scheduling is excluded since the locality-aware scheduler does not currently support querying numactl for page locality information. Locality-aware scheduling, in combination with heuristic-guided data distribution, improves or maintains performance compared to work-stealing.