0

I have METIS input file with edge weights of a graph. And I want to partition the graph with METIS, such that the edgecut in the METIS summary report is the least possible with METIS. It can be rb, kway or other algorithms or options.

So what options would work the best?

At the moment, following options have worked the best for me.

gpmetis -ptype=rb metis.input.file num.of.partitions

csrockstar
  • 55
  • 1
  • 9

1 Answers1

0

Using k-way partitioning (-ptype=kway) let's you choose to minimize the edgecut (-objtype=cut) or minimize the total communications volume (-objtype=vol). These two concepts are very similar (refer to METIS manual).

I found minimizing the total communication volume worked better than edgecut, This was because, when edges had weights, the edgecut in METIS is defined as the sum of edge weights cut. Minimizing this "edgecut" turned out to produce more border vertices.

Edgecut did perform a lot better if all the edge weights were (temporarily) set to 1, but it's simpler just to use the other option.

Source: METIS manual and personal experience.

ninjzz
  • 3
  • 5