[petsc-users] Streams scaling.log

Smith, Barry F. bsmith at mcs.anl.gov
Wed Apr 25 14:30:44 CDT 2018



> On Apr 25, 2018, at 2:12 PM, Manuel Valera <mvalera-w at sdsu.edu> wrote:
> 
> Hi and thanks for the quick answer,
> 
> Yes it looks i am using MPICH for my configure instead of using the system installation of OpenMPI, in the past i had better experience using MPICH but maybe this will be a conflict, should i reconfigure using the system MPI installation?
> 
> I solved the problem in a different way by login into the nodes i wanted to use and doing the make streams tests there, but i get the following:
> 
> np  speedup
> 1 1.0
> 2 1.51
> 3 2.17
> 4 2.66
> 5 2.87
> 6 3.06
> 7 3.44
> 8 3.84
> 9 3.81
> 10 3.17
> 11 3.69
> 12 3.81
> 13 3.26
> 14 3.51
> 15 3.61
> 16 3.81
> 17 3.8
> 18 3.64
> 19 3.48
> 20 4.01
> 
> So very modest scaling, this is about the same i get with my application, how can i make it work faster?

   You can't, the memory bandwidth is the limiting factor on this machine (not the number of cores) and there is nothing to be done about it. When buying machines make sure that the memory bandwidth is an important factor in the decision.

  Barry


> i am already using --map-by and --machinefile arguments for mpirun, maybe this is also a conflict with the different MPI installations?
> 
> Thanks,
> 
> 
> 
> On Wed, Apr 25, 2018 at 11:51 AM, Karl Rupp <rupp at iue.tuwien.ac.at> wrote:
> Hi Manuel,
> 
> this looks like the wrong MPI gets used. You should see an increasing number of processes, e.g.
> 
> Number of MPI processes 1 Processor names  node37
> Triad:         6052.3571   Rate (MB/s)
> Number of MPI processes 2 Processor names  node37 node37
> Triad:        9138.9376   Rate (MB/s)
> Number of MPI processes 3 Processor names  node37 node37 node37
> Triad:        11077.5905   Rate (MB/s)
> Number of MPI processes 4 Processor names  node37 node37 node37 node37
> Triad:        12055.9123   Rate (MB/s)
> 
> Best regards,
> Karli
> 
> 
> 
> 
> On 04/25/2018 08:26 PM, Manuel Valera wrote:
> Hi,
> 
> I'm running scaling tests on my system to check why my scaling is so poor, and after following the MPIVersion guidelines my scaling.log output looks like this:
> 
> Number of MPI processes 1 Processor names  node37
> Triad:        12856.9252   Rate (MB/s)
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Triad:         9138.3320   Rate (MB/s)
> Triad:         9945.0006   Rate (MB/s)
> Triad:        10480.8471   Rate (MB/s)
> Triad:        12055.4846   Rate (MB/s)
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Triad:         7394.1014   Rate (MB/s)
> Triad:         5528.9757   Rate (MB/s)
> Triad:         6052.7506   Rate (MB/s)
> Triad:         6188.5710   Rate (MB/s)
> Triad:         6944.4515   Rate (MB/s)
> Triad:         7407.1594   Rate (MB/s)
> Triad:         9508.1984   Rate (MB/s)
> Triad:        10699.7551   Rate (MB/s)
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Number of MPI processes 1 Processor names  node37
> Triad:         6682.3749   Rate (MB/s)
> Triad:         6825.3243   Rate (MB/s)
> Triad:         7217.8178   Rate (MB/s)
> Triad:         7525.1025   Rate (MB/s)
> Triad:         7882.1781   Rate (MB/s)
> Triad:         8071.1430   Rate (MB/s)
> Triad:        10341.9424   Rate (MB/s)
> Triad:        10418.4740   Rate (MB/s)
> 
> 
> Is this normal? i feel is different from what i get from an usual streams test, how can i get it to work properly?
> 
> Thanks,
> 
> 
> 
> 
> 



More information about the petsc-users mailing list