[petsc-users] Running problem with pc_type hypre

[Barry Smith]

  Some possibilities:

  Are you sure that the hypre was compiled with exactly the same MPI as the that used to build PETSc?

On May 28, 2014, at 4:57 PM, Danyang Su <danyang.su at gmail.com> wrote:

> Hi All,
> 
> I am testing my codes under windows with PETSc V3.4.4.
> 
> When running with option -pc_type hypre using 1 processor, the program exactly uses 6 processors (my computer is 6 processors 12 threads)

    6 threads? or 6 processes? It should not be possible for it to use more processes then what you start the program with.

    hypre can be configured to use OpenMP thread parallelism PLUS MPI parallelism. Was it configured/compiled for that? If so you want to turn that off,
configure and compile hypre before linking to PETSc so it does not use OpenMP.

   Are you sure you don’t have a bunch of zombie MPI processes running from previous jobs that crashed. They suck up CPU but are not involved in the current MPI run. Reboot the machine to get rid of them all.

  Barry

> and the program crashed after many timesteps. The error information is as follows:
> 
> job aborted:
> [ranks] message
> 
> [0] fatal error
> Fatal error in MPI_Comm_create: Internal MPI error!, error stack:
> MPI_Comm_create(536).......: MPI_Comm_create(comm=0x84000000, group=0xc80300f2, new_comm=0x000000001EA6DD30) failed
> MPI_Comm_create(524).......:
> MPIR_Comm_create_intra(209):
> MPIR_Get_contextid(253)....: Too many communicators
> 
> When running with option -pc_type hypre using 2 processors or more, the program exactly uses all the threads, making the system seriously overburden and the program runs very slowly.
> 
> When running without -pc_type hypre, the program works fine without any problem.
> 
> Does anybody have the same problem in windows.
> 
> Thanks and regards,
> 
> Danyang