Thomas : > >> I attached my data to this mail. For the largest matrix, umfpack failed > after allocating 4 GB of memory. I have not tried to figure out what's the > problem there. As you can see, for these matrices the distributed solvers > are
umfpack is a sequential package. 4GB+ likely exceeds memory capability of single core in your machine. slower by a factor of 2 or 3 compared to umfpack. For all solvers, I have > used the standard parameters, so I have not played around with the > permutation strategies and such things. This may be also the reason why > superlu is much slower than superlu_dist with just one core as it makes use > of different col and row permutation strategies. The data on superlu_dist and mumps look reasonable to me. The poor parallel performance is likely due to the multicore machine being used. Try these runs on a machine that is desirable for distributed computing. See http://www.mcs.anl.gov/petsc/documentation/faq.html#computers. Hong > > However, umpack will not work on a distributed memory machine. >> My personal preference is to use superlu_dist in parallel. In my >> experience using it as a coarse grid solver for multigrid, I find it >> much more reliable than mumps. However, when mumps works, its is >> typically slightly faster than superlu_dist. Again, not by a large >> amount - never more than a factor of 2 faster. >> > In my codes I also make use of the distributed direct solvers for the > coarse grid problems. I just wanted to make some tests how far away these > solvers are from the sequential counterparts. > > Thomas > > >> The failure rate using mumps is definitely higher (in my experience) >> when running on large numbers of cores compared to superlu_dist. I've >> never got to the bottom as to why it fails. >> >> Cheers, >> Dave >> >> >> On 15 May 2012 09:25, Thomas Witkowski<thomas.witkowski at >> tu-**dresden.de<thomas.witkowski at tu-dresden.de>> >> wrote: >> >>> I made some comparisons of using umfpack, superlu, superlu_dist and >>> mumps to >>> solve systems with sparse matrices arising from finite element method. >>> The >>> size of the matrices range from around 50000 to more than 3 million >>> unknowns. I used 1, 2, 4, 8 and 16 nodes to make the benchmark. Now, I >>> wonder that in all cases the sequential umfpack was the fastest one. So >>> even >>> with 16 cores, superlu_dist and mumps are slower. Can anybody of you >>> confirm >>> this observation? Are there any other parallel direct solvers around >>> which >>> are more efficient? >>> >>> Thomas >>> >> > -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20120515/243d1f4a/attachment.htm>