[Beowulf] How to debug error with Open MPI 3 / Mellanox / Red Hat?
Gus Correa
gus at ldeo.columbia.edu
Tue Apr 30 13:20:03 PDT 2019
It may be using IPoIB (TCP/IP over IB), not verbs/rdma.
You can force it to use openib (verbs, rdma) with (vader is for in-node
shared memory):
mpirun --mca btl openib,self,vader ...
These flags may also help tell which btl (byte transport layer) is being used:
--mca btl_base_verbose 30
See these FAQ:
https://www.open-mpi.org/faq/?category=openfabrics#ib-btl
https://www.open-mpi.org/faq/?category=all#tcp-routability-1.3
Better really ask more details in the Open MPI list. They are the pros!
My two cents,
Gus Correa
On Tue, Apr 30, 2019 at 3:57 PM Faraz Hussain <info at feacluster.com> wrote:
> Thanks, after buidling openmpi 4 from source, it now works! However it
> still gives this message below when I run openmpi with verbose setting:
>
> No OpenFabrics connection schemes reported that they were able to be
> used on a specific port. As such, the openib BTL (OpenFabrics
> support) will be disabled for this port.
>
> Local host: lustwzb34
> Local device: mlx4_0
> Local port: 1
> CPCs attempted: rdmacm, udcm
>
> However, the results from my latency and bandwith tests seem to be
> what I would expect from infiniband. See:
>
> [hussaif1 at lustwzb34 pt2pt]$ mpirun -v -np 2 -hostfile ./hostfile
> ./osu_latency
> # OSU MPI Latency Test v5.3.2
> # Size Latency (us)
> 0 1.87
> 1 1.88
> 2 1.93
> 4 1.92
> 8 1.93
> 16 1.95
> 32 1.93
> 64 2.08
> 128 2.61
> 256 2.72
> 512 2.93
> 1024 3.33
> 2048 3.81
> 4096 4.71
> 8192 6.68
> 16384 8.38
> 32768 12.13
> 65536 19.74
> 131072 35.08
> 262144 64.67
> 524288 122.11
> 1048576 236.69
> 2097152 465.97
> 4194304 926.31
>
> [hussaif1 at lustwzb34 pt2pt]$ mpirun -v -np 2 -hostfile ./hostfile ./osu_bw
> # OSU MPI Bandwidth Test v5.3.2
> # Size Bandwidth (MB/s)
> 1 3.09
> 2 6.35
> 4 12.77
> 8 26.01
> 16 51.31
> 32 103.08
> 64 197.89
> 128 362.00
> 256 676.28
> 512 1096.26
> 1024 1819.25
> 2048 2551.41
> 4096 3886.63
> 8192 3983.17
> 16384 4362.30
> 32768 4457.09
> 65536 4502.41
> 131072 4512.64
> 262144 4531.48
> 524288 4537.42
> 1048576 4510.69
> 2097152 4546.64
> 4194304 4565.12
>
> When I run ibv_devinfo I get:
>
> [hussaif1 at lustwzb34 pt2pt]$ ibv_devinfo
> hca_id: mlx4_0
> transport: InfiniBand (0)
> fw_ver: 2.36.5000
> node_guid: 480f:cfff:fff5:c6c0
> sys_image_guid: 480f:cfff:fff5:c6c3
> vendor_id: 0x02c9
> vendor_part_id: 4103
> hw_ver: 0x0
> board_id: HP_1360110017
> phys_port_cnt: 2
> Device ports:
> port: 1
> state: PORT_ACTIVE (4)
> max_mtu: 4096 (5)
> active_mtu: 1024 (3)
> sm_lid: 0
> port_lid: 0
> port_lmc: 0x00
> link_layer: Ethernet
>
> port: 2
> state: PORT_DOWN (1)
> max_mtu: 4096 (5)
> active_mtu: 1024 (3)
> sm_lid: 0
> port_lid: 0
> port_lmc: 0x00
> link_layer: Ethernet
>
> I will ask the openmpi mailing list if my results make sense?!
>
>
> Quoting Gus Correa <gus at ldeo.columbia.edu>:
>
> > Hi Faraz
> >
> > By all means, download the Open MPI tarball and build from source.
> > Otherwise there won't be support for IB (the CentOS Open MPI packages
> most
> > likely rely only on TCP/IP).
> >
> > Read their README file (it comes in the tarball), and take a careful look
> > at their (excellent) FAQ:
> > https://www.open-mpi.org/faq/
> > Many issues can be solved by just reading these two resources.
> >
> > If you hit more trouble, subscribe to the Open MPI mailing list, and ask
> > questions there,
> > because you will get advice directly from the Open MPI developers, and
> the
> > fix will come easy.
> > https://www.open-mpi.org/community/lists/ompi.php
> >
> > My two cents,
> > Gus Correa
> >
> > On Tue, Apr 30, 2019 at 3:07 PM Faraz Hussain <info at feacluster.com>
> wrote:
> >
> >> Thanks, yes I have installed those libraries. See below. Initially I
> >> installed the libraries via yum. But then I tried installing the rpms
> >> directly from Mellanox website (
> >> MLNX_OFED_LINUX-4.5-1.0.1.0-rhel7.5-x86_64.tar ). Even after doing
> >> that, I still got the same error with openmpi. I will try your
> >> suggestion of building openmpi from source next!
> >>
> >> root at lustwzb34:/root # yum list | grep ibverbs
> >> libibverbs.x86_64 41mlnx1-OFED.4.5.0.1.0.45101
> >> libibverbs-devel.x86_64 41mlnx1-OFED.4.5.0.1.0.45101
> >> libibverbs-devel-static.x86_64 41mlnx1-OFED.4.5.0.1.0.45101
> >> libibverbs-utils.x86_64 41mlnx1-OFED.4.5.0.1.0.45101
> >> libibverbs.i686 17.2-3.el7
> >> rhel-7-server-rpms
> >> libibverbs-devel.i686 1.2.1-1.el7
> >> rhel-7-server-rpms
> >>
> >> root at lustwzb34:/root # lsmod | grep ib
> >> ib_ucm 22602 0
> >> ib_ipoib 168425 0
> >> ib_cm 53141 3 rdma_cm,ib_ucm,ib_ipoib
> >> ib_umad 22093 0
> >> mlx5_ib 339961 0
> >> ib_uverbs 121821 3 mlx5_ib,ib_ucm,rdma_ucm
> >> mlx5_core 919178 2 mlx5_ib,mlx5_fpga_tools
> >> mlx4_ib 211747 0
> >> ib_core 294554 10
> >>
> >>
> rdma_cm,ib_cm,iw_cm,mlx4_ib,mlx5_ib,ib_ucm,ib_umad,ib_uverbs,rdma_ucm,ib_ipoib
> >> mlx4_core 360598 2 mlx4_en,mlx4_ib
> >> mlx_compat 29012 15
> >>
> >>
> rdma_cm,ib_cm,iw_cm,mlx4_en,mlx4_ib,mlx5_ib,mlx5_fpga_tools,ib_ucm,ib_core,ib_umad,ib_uverbs,mlx4_core,mlx5_core,rdma_ucm,ib_ipoib
> >> devlink 42368 4 mlx4_en,mlx4_ib,mlx4_core,mlx5_core
> >> libcrc32c 12644 3 xfs,nf_nat,nf_conntrack
> >> root at lustwzb34:/root #
> >>
> >>
> >>
> >> > Did you install libibverbs (and libibverbs-utils, for information and
> >> > troubleshooting)?
> >>
> >> > yum list |grep ibverbs
> >>
> >> > Are you loading the ib modules?
> >>
> >> > lsmod |grep ib
> >>
> >>
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://beowulf.org/pipermail/beowulf/attachments/20190430/413d6aef/attachment-0001.html>
More information about the Beowulf
mailing list