Fix ParMETIS call so it uses all workers in cluster.

[thvasilo]

Issue #, if available:

Description of changes:

• Previously we were calling the pm_dglpart executable through mpirun using only a single/process worker. This PR pulls in a necessary file that parmetis_preprocess.py creates in the wrong path, and fixes the number of workers used to match the number of partitions.

Detailed motivation for the change

The parmetis program implicitly expects a file named <graph-name>_stats.txt to be in the current working directory, the way the call is implemented at least. Actually the reference DGL implementation is wrong, it passes the graph_name whereas pm_dglpart actually expects a prefix path to which it appends _stats.txt. It just so happens that if we prefix the graph_name to a file in the cwd, and name it graph-name_stats.txt pm_dglpart will pick it up and use it.

A more correct solution would be to pass the absolute path to graph-name_stats.txt as the first argument to pm_dglpart (which pm_dglpart refers to as fstem)

You can check out the actual call here

The fstem var is used downstream here

Testing

Tested using Docker compose on ml-25M with 2 partitions.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[jalencato]

LGTM