Configuring the Hybrid Offload Version of the Intel Optimized MP LINPACK Benchmark

The most significant parameters in HPL.dat are N, NB, P and Q. Specify them as follows, as well as some other parameters:

• P and Q - the number of rows and columns in the process grid, respectively.

  P*Q must be the number of MPI processes that HPL is using.

  For the hybrid offload version of the Intel Optimized MP LINPACK Benchmark, keep P and Q roughly the same size.

  Note

  This setting is different from settings recommended for most HPL implementations, which usually recommend to choose P < Q and possibly with the 1:4 ratio.

• NB - block size of the data distribution.

  The table below shows recommended values of NB for 2^nd and 3^rd generation Intel® Core™ processors and for different numbers of Intel Xeon Phi coprocessors. The values may vary and depend on the PCI Express settings and performance of main memory.

  No coprocessors	1 coprocessor	2 coprocessors	3 coprocessors
  256	960	1024	1200

• N - the problem size:

  • For homogeneous runs, choose N divisible by NB*LCM(P,Q), where LCM is the least common multiple of the two numbers.

  • For heterogeneous runs, refer to Heterogeneous Intel Optimized MP LINPACK Benchmark.

  Note that increasing N usually increases performance, but the size of N is bounded by memory.

• Other parameters.

  To use Intel MKL BPUSH algorithm for horizontal broadcast, in line 23 of HPL.dat, set the BCASTs parameter to 6.

Enabling NUMA on your system and running an MPI process for each NUMA socket usually improves the hybrid offload performance. Refer to your BIOS settings for enabling NUMA on your system. You can use HPL_MIC_DEVICE and HPL_MIC_SHAREMODE environment variables to share the Intel Xeon Phi coprocessors among MPI processes (see Environment Variables for the Hybrid Offload for details). The scripts runme_intel64 and runme_intel64_dynamic set these environment variables for you for a given number of MPI ranks per node.