conquest/benchmarks/matrix_multiply/README.md

15 lines
821 B
Markdown

# Testing different matrix multiplication kernels
This system can be used for profiling different matrix multiplication kernels.
Those can be chosen with the `MULT_KERN` variable in `system.make`.
The additional coordinate files `si_XYZ.xtl` can be used to test weak scaling and
would work well for increasing the number of nodes: `si_222.xtl` is the same as `coords.dat`
and has 64 atoms. This means it would run well on anywhere from 2MPI/4OpenMP to 8MPI/1OpenMP.
With the rest of the `xtl` files, we double the number of atoms each time, and would need
to double the number of processes.
We now have systems from 64 atoms (222) to 262144 atoms (323232) which will scale from 8 MPI
to 32,768 MPI processes (1 OpenMP thread) with 8 atoms per process or 1 MPI process to 4096 MPI
processes with 64 atoms per process.