Re: [petsc-users] Replicating a hang with MUMPS

2018-06-27 Thread David Knezevic
On Wed, Jun 27, 2018 at 3:12 PM, Smith, Barry F. wrote: > > David, > > This is ugly but should work. BEFORE reading in the matrix and right > hand side set the LOCAL sizes for the matrix and vector. This way you can > control exactly which rows go on which process. Note you will have to ha

Re: [petsc-users] Replicating a hang with MUMPS

2018-06-27 Thread Smith, Barry F.
David, This is ugly but should work. BEFORE reading in the matrix and right hand side set the LOCAL sizes for the matrix and vector. This way you can control exactly which rows go on which process. Note you will have to have your own mechanism to know what the local sizes should be (fo

Re: [petsc-users] Replicating a hang with MUMPS

2018-06-27 Thread Stefano Zampini
Forwarding to the list Il Mer 27 Giu 2018, 21:46 Stefano Zampini ha scritto: > You can call MatSetSizes before MatLoad > > Il Mer 27 Giu 2018, 21:36 David Knezevic ha > scritto: > >> I ran into a case where using MUMPS (called via "-ksp_type preonly >> -pc_type lu -pc_factor_mat_solver_package

[petsc-users] Replicating a hang with MUMPS

2018-06-27 Thread David Knezevic
I ran into a case where using MUMPS (called via "-ksp_type preonly -pc_type lu -pc_factor_mat_solver_package mumps") for a particular solve hangs indefinitely with 24 MPI processes (but it works fine with other numbers of processes). The stack trace when killing the job is below, in case that gives