Re: [OMPI users] MPI_Finalize not behaving correctly, orphaned processes

2015-04-24 Thread Jeff Squyres (jsquyres)
Does the child process end up with valid memory in the buffer in that sample? Back when I paid attention to verbs (which was admittedly a long time ago), the sample I pasted would segv... > On Apr 24, 2015, at 9:40 AM, Mike Dubman wrote: > > ibv_fork_init() will set special flag for madvise(

Re: [OMPI users] MPI_Finalize not behaving correctly, orphaned processes

2015-04-24 Thread Mike Dubman
ibv_fork_init() will set special flag for madvise() (IBV_DONTFORK/DOFORK) to inherit (and not cow) registered/locked pages on fork() and will maintain refcount for cleanup. I think some minimal kernel version required (2.6.x) which supports these flags. I can check if internally if you think the

Re: [OMPI users] Hang in MPI_Comm_split in 2 RHEL Linux nodes with INTEL MIC cards

2015-04-24 Thread George Bosilca
Manumachu, Both nodes have the same IP for their Phi (mic0 and mic1). This is OK as long as they don't try to connect to each other using these addresses. A simple fix is to prevent OMPI from using the supposedly local mic0 and mic1 IP. Add --mca btl_tcp_if_exclude mic0,mic1 to your mpirun comm

Re: [OMPI users] MPI_Finalize not behaving correctly, orphaned processes

2015-04-24 Thread Jeff Squyres (jsquyres)
Mike -- What happens when you do this? ibv_fork_init(); int *buffer = malloc(...); ibv_reg_mr(buffer, ...); if (fork() != 0) { // in the child *buffer = 3; // ... } > On Apr 24, 2015, at 2:54 AM, Mike Dubman wrote: > > btw, ompi master now calls ibv_fork_init() before

Re: [OMPI users] MPI_THREAD_MULTIPLE and openib btl

2015-04-24 Thread Mike Dubman
yes #1 - ob1 as pml, openib openib as btl (default: rc) #2 - yalla as pml, mxm as IB library (default: ud, use "-x MXM_TLS=rc,self,shm" for rc) #3 - cm as pml, mxm as mtl and mxm as a transport (default: ud, use params from #2 for rc) On Fri, Apr 24, 2015 at 10:46 AM, Subhra Mazumdar wrote: > I

Re: [OMPI users] MPI_THREAD_MULTIPLE and openib btl

2015-04-24 Thread Subhra Mazumdar
I am a little confused now, I ran 3 different ways and got 3 different performance from best to worse in following order: 1) mpirun --allow-run-as-root --mca pml ob1 -n 1 /root/backend localhost : -x LD_PRELOAD=/root/libci.so -n 1 /root/app2 2) mpirun --allow-run-as-root -n 1 /root/backend lo

[OMPI users] Hang in MPI_Comm_split in 2 RHEL Linux nodes with INTEL MIC cards

2015-04-24 Thread Manumachu Reddy
Dear OpenMPI Users, I request your help to resolve a hang in my OpenMPI application. My OpenMPI application hangs in MPI_Comm_split() operation. The code for this simple application is at the end of this email. Broadcast works fine. My experimental setup comprises of two RHEL6.4 Linux nodes. Eac

Re: [OMPI users] MPI_THREAD_MULTIPLE and openib btl

2015-04-24 Thread Mike Dubman
HPCX package uses pml "yalla" by default (part of ompi master branch, not in v1.8). So, "-mca mtl mxm" has no effect, unless "-mca pml cm" specified to disable "pml yalla" and let mtl layer to play. On Fri, Apr 24, 2015 at 6:36 AM, Subhra Mazumdar wrote: > I changed my downloaded MOFED versio

Re: [OMPI users] MPI_Finalize not behaving correctly, orphaned processes

2015-04-24 Thread Mike Dubman
btw, ompi master now calls ibv_fork_init() before initializing btl/mtl/oob frameworks and all fork fears should be addressed. On Fri, Apr 24, 2015 at 4:37 AM, Jeff Squyres (jsquyres) wrote: > Disable the memory manager / don't use leave pinned. Then you can > fork/exec without fear (because on