In addition to what Wolfgang wrote, Sometimes the OS will map the cores of a CPU in the following order: Physical-Logical-Physical-Logical-Physical-Logical etc. Consequently, if your processor supports hyperthreading through the use of logical core, running with core 0-1 means you are essentially running on a single physical core but using two logical cores. For intensive operations that use the same instruction set, this is generally never a good idea. In the present case though, since the assembly is fast but not the matrix solve, I would presume it is more a question of memory lane issues than anything else. Your CPU might be really memory-bandwith bound.
On Wednesday, November 15, 2023 at 11:39:59 a.m. UTC-5 Wolfgang Bangerth wrote: > > > On 11/15/23 04:09, Abbas Ballout wrote: > > > > If I run mpirun -np2 the assemble times are 25.8 seconds and the MUMPS > > solve times are 51.7 seconds. > > If I run mpirun --cpu-set 0-1 -np 2, the assemble times are 26 seconds > > (unchanged) but the solve time are at 94.9 seconds! > > Is this normal and expected? > > What happens if you use `--cpu-set 0,2` or other combinations? > > Cores have memory channels that they often share with other cores. I > wonder whether, for example, cores 0 and 1 share a memory channel and > consequently step on each other's toes during the direct solve (which is > memory bound). Or they share SIMD execution units. It would be > interesting to see what happens if you choose other sets of cores to use. > > Best > W. > -- The deal.II project is located at http://www.dealii.org/ For mailing list/forum options, see https://groups.google.com/d/forum/dealii?hl=en --- You received this message because you are subscribed to the Google Groups "deal.II User Group" group. To unsubscribe from this group and stop receiving emails from it, send an email to dealii+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/dealii/d4dfee09-df51-4f87-90ad-71b1cd7eea52n%40googlegroups.com.
