Parallel Solver is Slow

Linear Algebra
1

Hello,

I’m trying to speed up the time of my solver by solving in parallel. I’ve tried this for the demo ft01_poisson.py with a mesh size of 1000 X 1000-- I ran the command:

mpirun -n 10 python3 ft01_poisson.py

However, this takes longer than running the demo in the usual way:

python3 ft01_poisson.py

As I understand, the “-n 10” flag, splits the domain into 10 pieces, solves them separately, and patches the solutions together somehow. Is there something else I need to do in order to tell my computer to use multiple threads?

In a similar post, it was suggested to set environment variable: set OPT_NUM_THREADS = 1. However, it is still slow after I do this.

I’m new to parallel processing so any help is greatly appreciated.

Thanks,
Trevor

2

Just to be sure, how many processors do you have available? What does nproc report?

3

The ‘nproc’ command isn’t available on macs, so I did instead:

sysctl -n hw.ncpu

This returns 4.

4

It means you’re oversubscribing your CPUs with 2-3 processes each, so you would expect it to run slow. Keep your mpi to the number of processors available.

5

Be careful regarding the difference between processes and threads.

6

Should be OMP_NUM_THREADS.

1 Like