Getting High Performance on NUMA-based Nehalem-EX System with MKL Without Controlling NUMA Policy for Shared Memory


Traditional MKL performance tests being executed on Nehalem-EX by default got rather poor performance on a number of MKL LAPACK functions, particularly on DGETRF. It's been detected that using such utility as numactll on that system under Linux, namely controlling NUMA policy for shared memory, improves performance dramatically , that is, starting an application in this way:

  • Intel® Math Kernel Library
  • Nehalem-EX
  • Michael Chuvelev
  • Numa Policy for Shared Memory
  • MLK
  • LAPACK functions
  • numactl
  • Parallel Computing
  • Subscribe to MLK