Quality of Matrix Ordering of ParMETIS
I test a parallel sparse solver MUMPS using ParMETIS. For a relevant small sparse matrix(N=52116,NZ=288704), if I use 1-7 processors, MUMPS works ok. But if use 8 processor, MUMPS gives error.
Q1: Does ParMETIS have any issue regarding using too many processors?
I also found in general, METIS produces better ordering than ParMETIS.
Q2: Is it true that METIS usually generate better ordering than ParMETIS?
For a relevant large matrix(N=5389726,NZ=42250735), METIS crashes due to insufficient memory on my local host.
Q3: Does ParMETIS help in this case if I use cluster? In other words, does ParMETIS use distributed memory or not?
Submitted by sam.guo@us.cd-a... on Fri, 2011-02-04 13:46
»
- Login to post comments
RE: Sam, You should try the newly
Sam,
You should try the newly posted 3.2 version of ParMetis. It has much improved fill-reducing ordering routines.
George