Thanks again, Hui. It looks like I hadn't understood the structure or delegation of responsibility between mpich and libfabric, nor the extent to which libfabric itself comprised part of the runtime interface. I'll give this a try, and I expect this knowledge will be useful to me for other MPI that use libfabric as well! ~jonathon On Fri, Jul 1, 2022 at 7:46 AM Zhou, Hui <[email protected]> wrote:

For libfabric, which used by ch4:ofi, you can these settings from libfabric manpages. It depends on your provider. For example, https://ofiwg.github.io/libfabric/v1.11.1/man/fi_tcp.7.html fi_tcp(7) - GitHub Pages <https://ofiwg.github.io/libfabric/v1.11.1/man/fi_tcp.7.html> NAME. fi_tcp - The msg sockets Fabric Provider. OVERVIEW. The tcp provider can be used on any system that supports TCP sockets. The provider is not intended to provide performance improvements over regular TCP sockets, but rather to allow developers to write, test,and debug application code even on platforms that do not have high-performance fabric hardware. ofiwg.github.io

------------------------------ *From:* Jonathon Anderson <[email protected]> *Sent:* Friday, July 1, 2022 8:40 AM *To:* Zhou, Hui <[email protected]> *Cc:* [email protected] <[email protected]> *Subject:* Re: [mpich-discuss] How to specify which Ethernet / IP interface to use?

Thank you very much for this information, Hi! Can you point me to where these environment variables are documented for future reference? I tried to find something like this and was unable.

~jonathon

On Fri, Jul 1, 2022, 7:35 AM Zhou, Hui <[email protected]> wrote:

Hi Jonathon,

Since the binary name is "mpich-ofi-gnu9-ohpc-3.4.2-3.1.ohpc.2.4.x86_64", I assume it is mpich -3.4.2 configured with ch4:ofi. We no longer use mpich2 to refer to the project, just mpich.

With ch4:ofi, try setting the following environment variables: FI_PROVIDER=tcp FI_TCP_IFACE=eth0

That assumes you want to use the tcp provider and the private interface is eth0. There are other providers that you may choose and the iface selection variables will be different. You can set MPIR_CVAR_DEBUG_SUMMARY=1 to show the list of providers mpich probes during init.

Best wishes, -- Hui ------------------------------ *From:* Jonathon Anderson via discuss <[email protected]> *Sent:* Thursday, June 30, 2022 10:49 PM *To:* [email protected] <[email protected]> *Cc:* Jonathon Anderson <[email protected]> *Subject:* Re: [mpich-discuss] How to specify which Ethernet / IP interface to use?

I'm trying to use mpich2 [1] on a cluster of compute nodes each of which has two Ethernet interfaces: one, an encumbered public / Internet interface, and another a private / intranet interface. mpich is defaulting to trying to use the public interface, which then hangs indefinitely with both nodes in my job stuck listening via UDP.

I experience the same behavior whether I use mpiexec or srun --mpi=pmi2.

How do I tell mpich which interface to use?

Thanks!

~jonathon

[1]: mpich-ofi-gnu9-ohpc-3.4.2-3.1.ohpc.2.4.x86_64 _______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss