Hi Mohammed, I don’t recommend to verbs provider with FI_EP_RDM EP type. We deprecating this provider since libfabric 1.7. Now our default provider is RxM/verbs (verbs with FI_EP_MSG EP type) to run libfabric on IB/iWARP/RoCE verbs devices. If you use libfabric 1.6.x, RxM/verbs should be the default provider.
Can you see the hangs when using libfabric’s master branch? -- Dmitry From: Mohammed Shaheen [mailto:[email protected]] Sent: Thursday, November 22, 2018 4:54 PM To: Gladkov, Dmitry <[email protected]>; Hefty, Sean <[email protected]>; [email protected]; [email protected]; Ilango, Arun <[email protected]> Subject: Re: [libfabric-users] intel mpi with libfabric Thanks Arun and Dmitry for your support. Well, I am building my own libfabric, and I export the right variables and source intel mpi with -ofi_internal=0. I figured out where the problem is: 1. If libfabric is built for all providers, i.e. run ./configure without including and exluding providers, it will build ibverbs among others; however, the mpi test program will hang during execution. 2. If libfabric configured with only enabling ibverbs and setting all other providers, i.e. ./configure --enable-verbs=yes --enable-rxm=no --enable-rxd=no --enable-sockets=no --enable-tcp=no --enable-udp=no, mpi test program will run through Another observation when I enable debug, --enable-debug, I get the aforementioned message (here it is again): prov/verbs/src/ep_rdm/verbs_rdm_cm.c:337: fi_ibv_rdm_process_addr_resolved: Assertion `id->verbs == ep->domain->verbs' failed. and the mpi test program runs through in case 2 above. I am not sure whether or not I should take this message seriously? I did not see any difference in the test mpi program behaviour if I build ibverbs as a DSO (--enable-verbs=dl) or as the default which I suppose would be part of libfabric (--enable-verbs=yes) except in case of DSO, the FI_PROVIDER_PATH must be exported. However, worth mentioning as a bug (probably), when ibverbs (or any other provider I assume) is built as a DSO, the libfabric folder under which the provider DSOs are put has wrong permissions, which means if you build libfabric as a root and use default installation folders (/usr/local/lib), your mpi program would not run through if you launch it as some other user. Regards, Mohammed Am Mittwoch, 21. November 2018, 19:42:24 MEZ hat Ilango, Arun <[email protected]<mailto:[email protected]>> Folgendes geschrieben: Mohammed, Just to add what Dmitry said, if you're using your own libfabric, please make sure it's the latest (i.e. v1.6.2). You can check the version by running fi_info --version. Other things to check: 1. Make sure you have librdmacm package installed. 2. Check if the IPoIB interface of the node has been configured with an IP address and is pingable from other nodes in the cluster. Thanks, Arun. -----Original Message----- From: Gladkov, Dmitry Sent: Wednesday, November 21, 2018 10:31 AM To: Hefty, Sean <[email protected]<mailto:[email protected]>>; Mohammed Shaheen <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>; [email protected]<mailto:[email protected]> Cc: Ilango, Arun <[email protected]<mailto:[email protected]>> Subject: RE: [libfabric-users] intel mpi with libfabric Hi Mohammed, Do you use your own version of libfabirc? IMPI 2019 U1 uses its internal libfabric by default. If you use your libfabric, please, specify LD_LIBRABRY_PATH to your library and FI_PROVIDER_PATH to path to OFI DL providers (<ofi_install_dir>/lib/libfabric) if you use DL provider, or unset this variable (mpivars.sh sets it). -- Dmitry -----Original Message----- From: Hefty, Sean Sent: Wednesday, November 21, 2018 8:52 PM To: Mohammed Shaheen <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>; [email protected]<mailto:[email protected]> Cc: Ilango, Arun <[email protected]<mailto:[email protected]>>; Gladkov, Dmitry <[email protected]<mailto:[email protected]>> Subject: RE: [libfabric-users] intel mpi with libfabric Copying ofiwg and key developers for this issue. - Sean > I get the following error running a small mpi test program using intel > mpi 2019 from intel parallel studio cluster edition update 1 (the > newest) on Mellanox FDR Cluster: > > > > test.e: prov/verbs/src/ep_rdm/verbs_rdm_cm.c:337: > fi_ibv_rdm_process_addr_resolved: Assertion `id->verbs == ep->domain- > >verbs' failed. > > > > The program hangs on this error message. I installed the newest > release of libfabric and configured it with only ibverbs support. I > used the inbox (sles 11 sp4 and sles 12 sp3) ibverbs and rdma > libraries. I also tried with mellanox ofed to no avail. > > > > > Any ideas how to go about it? > > > > > > Regards, > > Mohammed -------------------------------------------------------------------- Joint Stock Company Intel A/O Registered legal address: Krylatsky Hills Business Park, 17 Krylatskaya Str., Bldg 4, Moscow 121614, Russian Federation This e-mail and any attachments may contain confidential material for the sole use of the intended recipient(s). Any review or distribution by others is strictly prohibited. If you are not the intended recipient, please contact the sender and delete all copies.
_______________________________________________ ofiwg mailing list [email protected] https://lists.openfabrics.org/mailman/listinfo/ofiwg
