Re: [OMPI users] Open MPI 1.10.0: Works on one Sandybridge Node, not on another: tcp_peer_send_blocking

Try adding —novm to that cmd line - if all the nodes are topologically identical (minus the GPU), then this will speed things up. > On Sep 24, 2015, at 9:26 AM, Matt Thompson wrote: > > On Thu, Sep 24, 2015 at 12:10 PM, Ralph Castain

Re: [OMPI users] Open MPI 1.10.0: Works on one Sandybridge Node, not on another: tcp_peer_send_blocking

On Thu, Sep 24, 2015 at 12:10 PM, Ralph Castain wrote: > Ah, sorry - wrong param. It’s the out-of-band that is having the problem. > Try adding —mca oob_tcp_if_include > Ooh. Okay. Look at this: (13) $ mpirun --mca oob_tcp_if_include ib0 -np 2 ./helloWorld.x Process 1 of 2

Re: [OMPI users] Open MPI 1.10.0: Works on one Sandybridge Node, not on another: tcp_peer_send_blocking

Ah, sorry - wrong param. It’s the out-of-band that is having the problem. Try adding —mca oob_tcp_if_include > On Sep 24, 2015, at 8:56 AM, Matt Thompson wrote: > > Ralph, > > I believe these nodes might have both an Ethernet and Infiniband port where > the Ethernet port

Re: [OMPI users] Open MPI 1.10.0: Works on one Sandybridge Node, not on another: tcp_peer_send_blocking

Ralph, I believe these nodes might have both an Ethernet and Infiniband port where the Ethernet port is not the one to use. Is there a way to tell Open MPI to ignore any ethernet devices it sees? I've tried: --mca btl sm,openib,self and (based on the advice of the much more intelligent support

Re: [OMPI users] Open MPI 1.10.0: Works on one Sandybridge Node, not on another: tcp_peer_send_blocking

Starting in the 1.7 series, OMPI by default launches daemons on all nodes in the allocation during startup. This is done so we can “probe” the topology of the nodes and use that info during the process mapping procedure - e.g., if you want to map-by NUMA regions. What is happening here is that

[OMPI users] Open MPI 1.10.0: Works on one Sandybridge Node, not on another: tcp_peer_send_blocking

Open MPI Users, I'm hoping someone here can help. I built Open MPI 1.10.0 with PGI 15.7 using this configure string: ./configure --disable-vt --with-tm=/PBS --with-verbs --disable-wrapper-rpath \ CC=pgcc CXX=pgCC FC=pgf90 F77=pgf77 CFLAGS='-fpic -m64' \ CXXFLAGS='-fpic -m64'

Re: [OMPI users] Problem using Open MPI 1.10.0 built with Intel compilers 16.0.0

Intel apparently changed something in their 2016 compiler (compared to the 2015 compiler); the Open MPI configure script decided to use a different pragma. Per the issue I opened up on Github, I need to look at the configure script and see what's going wrong. > On Sep 24, 2015, at 4:51 PM,

Re: [OMPI users] Problem using Open MPI 1.10.0 built with Intel compilers 16.0.0

Hi, I have made some other tests. I don't know if it can help you but here is what I observed. Using the array contructor [] solves the problem for a scalar, as someone wrote on the Intel forum. The same code with tok declared as an integer and call

Re: [OMPI users] Problem using Open MPI 1.10.0 built with Intel compilers 16.0.0

Yes -- typo -- it's not a problem with mpi_f08, it's a problem with the mpi module using the "ignore TKR" implementation. See https://github.com/open-mpi/ompi/issues/937. > On Sep 24, 2015, at 4:30 PM, Gilles Gouaillardet > wrote: > > Jeff, > > I am not sure

Re: [OMPI users] Problem using Open MPI 1.10.0 built with Intel compilers 16.0.0

BTW, I created this Github issue to track the problem: https://github.com/open-mpi/ompi/issues/937 > On Sep 24, 2015, at 4:27 PM, Jeff Squyres (jsquyres) > wrote: > > I looked into the MPI_BCAST problem -- I think we (Open MPI) have a problem > with the mpi_f08

Re: [OMPI users] Problem using Open MPI 1.10.0 built with Intel compilers 16.0.0

Jeff, I am not sure whether you made a typo or not ... the issue only occuex with f90 bindings (aka use mpi) f08 bindings (aka use mpi_f08) works fine Cheers, Gilles On Thursday, September 24, 2015, Jeff Squyres (jsquyres) wrote: > I looked into the MPI_BCAST problem --

Re: [OMPI users] Problem using Open MPI 1.10.0 built with Intel compilers 16.0.0

I looked into the MPI_BCAST problem -- I think we (Open MPI) have a problem with the mpi_f08 bindings and the Intel 2016 compilers. It looks like configure is choosing to generate a different pragma for Intel 2016 vs. Intel 2015 compilers, and that's causing a problem. Let me look into this a

Re: [OMPI users] OpenMPI-1.10.0 bind-to core error

Ha! I finally tracked it down - a new code path that bypassed the prior error output. I have a fix going into master shortly, and will then port it to 1.10.1. Thanks for your patience! Ralph > On Sep 24, 2015, at 1:12 AM, Patrick Begou > wrote: > > Sorry

Re: [OMPI users] Problem with implementation of Foxa algorithm

Oh... Thanks for fast answer! Now I understand where error. At least now code is work! But matrix calculated wrong... Now I need to think about my Fox algorithm implementation. 2015-09-24 3:54 GMT+03:00 Gilles Gouaillardet : > Hi, > > at line 171 > MPI_Gather([i*matrixSize],

Re: [OMPI users] Problem using Open MPI 1.10.0 built with Intel compilers 16.0.0

Hello, Thanks for the quick answer. I think I cannot use mpi_f08 in my code because I am also using parallel HDF5 which does not seem to be compatible with the Fortran 2008 module. I will ask Intel what they think about this problem. Thanks, Fabrice Le 24/09/2015 02:18, Gilles Gouaillardet

Re: [OMPI users] OpenMPI-1.10.0 bind-to core error

Sorry for the delay. Runing mpirun whith wrong OMPI_MCA_plm_rsh_agent doesn't give any explicit message in OpenMPI-1.10.0. How I can show the problem: I request 2 nodes, 1cpu on each node, 4 cores on each cpu (so 8 cores availables with cpusets). Node file is: [begou@frog7 MPI_TESTS]$ cat