Locality-aware task scheduling and data distribution on NUMA systems

Abstract

Modern parallel computer systems exhibit Non-Uniform Memory Access (NUMA) behavior. For best performance, any parallel program therefore has to match data allocation and scheduling of computations to the memory architecture of the machine. When done manually, this becomes a tedious process and since each individual system has its own peculiarities this also leads to programs that are not performance-portable. We propose the use of a data distribution scheme in which NUMA hardware peculiarities are abstracted away from the programmer and data distribution is delegated to a runtime system which is generated once for each machine. In addition we propose using task data dependence information now possible with the OpenMP 4.0RC2 proposal to guide the scheduling of OpenMP tasks to further reduce data stall times. We demonstrate the viability and performance of our proposals on a four socket AMD Opteron machine with eight NUMA nodes. We identify that both data distribution and locality-aware task scheduling improves performance compared to default policies while still providing an architecture-oblivious approach for the programmer. © 2013 Springer-Verlag.