* Both ch4:ucx,ofi and

ch4:ofi,ucx build
            with both the UCX and Libfabric, and one can use the
            environment variable "MPIR_CVAR_CH4_NETMOD" to choose which
            netmod (ofi or ucx) at runtime.

The order between ucx an ofi determines the default.

On single mode, shared memory will be used unless it is disabled. "ch4:ucx" disables shared memory by default since UCX's shared memory performance is quite decent. When shared memory is used, which FI_PROVIDER doesn't matter.

Hope this answers your question.

Cheers,

Hui

From: John Cary <cary@colorado.edu>
Sent: Monday, June 22, 2026 2:21 PM
To: discuss@mpich.org <discuss@mpich.org>; Zhou, Hui <zhouh@anl.gov>
Subject: Re: [mpich-discuss] How to get srun/mpich to use the right interface?

This Message Is From an External Sender

This message came from outside your organization.

Hello Hui,

We are getting some mysterious results. We have four configurations:

--with-device=ch4:ucx
--with-device=ch4:ofi
--with-device=ch4:ucx,ofi
--with-device=ch4:ofi,ucx

Is there any difference between the last two? Are the last two valid configurations?

If one is on a single node, does FI_PROVIDER=verbs make any sense?
Or perhaps on a single node, shared memory is always used?

Thanks....John

On 6/11/26 10:33 AM, John Cary wrote:

Thanks, Hui!

I am trying to have one build that works for our cluster and also for AWS, which uses ofi. Is that possible?

Thx....John

On 6/11/26 10:06 AM, Zhou, Hui wrote:
[External email - use caution]
HI John,

Try using the UCX instead of libfabric. You can configure MPICH with ./configure --with-device=ch4:ucx to use UCX. With libfabric, could you try set environment variable FI_PROVIDER=verbs ?

If you still have issue, try run a dummy MPI program setting `MPIR_CVAR_DEBUG_SUMMARY=1`. That will provide more logging details on which libfabric provider is being selected.

--

Hui Zhou

From: John Cary via discuss <discuss@mpich.org>
Sent: Thursday, June 11, 2026 8:50 AM
To: discuss@mpich.org <discuss@mpich.org>
Cc: John Cary <cary@colorado.edu>
Subject: [mpich-discuss] How to get srun/mpich to use the right interface?
How to get mpich to use the right interface? mpich configured and built with libfabric as shown below. It is run using slurm (srun). The result is [1781142435. 559314543] ne07: rank64. vorpal: Failed to modify UD QP to INIT on mlx5_0: Operation

ZjQcmQRYFpfptBannerStart

This Message Is From an External Sender

This message came from outside your organization.

ZjQcmQRYFpfptBannerEnd
How to get mpich to use the right interface?

mpich configured and built with libfabric as shown below.  It is run 
using slurm (srun).  The result is

[1781142435.559314543] ne07:rank64.vorpal: Failed to modify UD QP to 
INIT on mlx5_0: Operation not permitted
[1781142435.563376291] ne07:rank66.vorpal: Failed to modify UD QP to 
INIT on mlx5_0: Operation not permitted
Abort(203572367): Fatal error in internal_Init: Other MPI error, error stack
...

which I think means that mpich is txrying to run over the mlx5_0 
interface, which does not exist.  The interfaces are

eno1: flags=4099<UP,BROADCAST,MULTICAST>  mtu 1500
eno2: flags=4163<UP,BROADCAST,RUNNING,MULTICAST>  mtu 1500
ib0: flags=4163<UP,BROADCAST,RUNNING,MULTICAST>  mtu 2044
lo: flags=73<UP,LOOPBACK,RUNNING>  mtu 65536

and we want to use ib0, infiniband.  We tried

export HYDRA_IFACE=ib0
srun ...

and got the same error.

How can srun/mpich be instructed to use the ib0 interface by default?

Also, how can one see which interface mpich is choosing?

Thx...



'/user/builds-linux-centos8-zen2/gvxsimall/mpich-5.0.1/configure' \
--prefix=/home/research/user/installs/linux-centos8-zen2/contrib-gcc1140/mpich-5.0.1-shared 
\
   --enable-shared \
   --disable-static \
   CC='/home/common/gcc-11.4.0/bin/gcc' \
   CXX='/home/common/gcc-11.4.0/bin/g++' \
   FC='/home/common/gcc-11.4.0/bin/gfortran' \
   F77='/home/common/gcc-11.4.0/bin/gfortran' \
   CFLAGS='-pthread -pipe -fPIC' \
   CXXFLAGS='-pthread -Wno-deprecated-declarations -pipe -fPIC' \
   FFLAGS='-fallow-argument-mismatch -pipe -fPIC' \
   FCFLAGS='-fallow-argument-mismatch -pipe -fPIC' \
   LDFLAGS='-L/home/common/gcc-11.4.0/lib64 
-Wl,-rpath,/home/common/gcc-11.4.0/lib64' \
   LDSHARED='-L/home/common/gcc-11.4.0/lib64 
-Wl,-rpath,/home/common/gcc-11.4.0/lib64' \
   --with-libfabric=install \
   --with-device=ch4:ofi \
   --without-cuda \
   --disable-gl