On Mon, 23 Mar 2015, Lisandro Dalcin wrote: > On 23 March 2015 at 18:27, Satish Balay <[email protected]> wrote: > > nemesis has this performance degradation issue [when oversubscribing > > cpu-cores] - so --download-mpich defaults using > > -download-mpich-device=ch3:sock [this mode doesn't not have > > performance degradation] > > > > Well, the thing is I always use MPI from my system package manager > (yum or brew) and they are configured with nemesis.
yup - and thats a problem. http://trac.mpich.org/projects/mpich/ticket/79 I don't know if openmpi also has this issue or not. So I use a prebuilt mpich [wth ch3:sock] as my default MPI on my laptop. What you want is to tell PETSc to not run tests that require more than - say 8 nodes - but this feature is missing.. Satish
