Task Parallel Models Based on Dynamic Data Placement toReduce NUMA Effects

NUMA (Non-Uniform Memory Access) multicore computersbecome popular in scientific and industrial fields due to itsscalable memory performance. However, large-scale intensivedata computing on NUMA architecture are facing up to thechallenges in data locality problems called NUMA effects thatare caused by the overhead accesses of cross-node data. Our taskparallel model bases on the strategy of dynamic data placementimproving system performance by reducing the frequently dataaccess to remote memory and also by keeping load balancebetween each NUMA domain. The task parallel modelsinvolved OpenMP, numactl and libnuma. The evaluationdemonstrates that the benchmarks using our task parallemodels on a 32-core NUMA computer with various workloadsachieve system performance improvement by 50% at least.

کلیدواژه ها:

NUMA Architecture ، Task Parallel Model ، Dynamic Data Placement ، NUMA Effects