Intel® Math Kernel Library 2018 Update 1 Developer Guide
The Intel Distribution for LINPACK Benchmark supports command-line parameters for HPL that help you to avoid making small changes in the HPL.dat input file every time you do a new run.
Placeholders in this command line illustrate these parameters:
xhpl.exe -n <problem size> -m <memory size in Mbytes> -b <block size> -p <grid row dimn> -q <grid column dimn>
You can also use command-line parameters with the sample runme script. For example:
runme_intel64_dynamic.bat -m <memory size in Mbytes> -b <block size> -p <grid row dimn> -q <grid column dimn>
For more command-line parameters, see Heterogeneous Support in the Intel Distribution for LINPACK Benchmark.
If you want to run for N=10000 on a 1x3 grid, execute this command, provided that the other parameters in HPL.dat and the script are correct:
runme_intel64_dynamic.bat -n 10000 -p 1 -q 3
By using the m parameter you can scale by the memory size instead of the problem size. The m parameter only refers to the size of the matrix storage and not to the coprocessor memory size or other buffers. So if you want to use matrices that fit in 50000 Mbytes with NB=256 on 16 nodes, adjust the script to set the total number of MPI processes to 16 and execute this command:
runme_intel64_dynamic.bat -m 50000 -b 256 -p 4 -q 4