[ofiwg] noob questions

Jeff Squyres (jsquyres) jsquyres at cisco.com
Wed Nov 13 16:48:43 PST 2019

Just to clarify the Open MPI behavior for you...

PML = point to point messaging layer. PML plug-ins effect MPI point to point calls such as MPI_Send and MPI_Recv. 

Open MPI has a few different PMLs. One of them is CM, which is a super thin layer of high-level translation glue between Open MPI and backend APIs that can handle network level matching (such as various Libfabric providers). These back-end APIs are implemented in MTL (matching transport layer) plugins. MTLs are the low-level translation glue to the back-end matching-capable network API. Open MPI has an OFI MTL which makes Libfabric API calls. 

Hence: MPI_Send —> CM PML —> OFI MTL —> Libfabric. 

The OFI MTL expects to be able to use LibFabric RDM endpoints (you can see all the attributes it asked for in the log). 

You can pass CLI options to mpirun to effect which PML and MTL plugins are used (vs letting OMPI auto selecting). You can also pass in a variety of run time params to each of those plugins. 

If you tell Open MPI to only use CM and the OFI MTL but the OFI MTL fails to enable itself because it can’t find a provider that has all the requirements it’s looking for, that will result in CM disabling itself (because it can’t find an appropriate MTL to use). Then you get the error message you saw: peers can’t reach other. 

If you let OMPI auto select, CM will silently fail and a different PML may be selected, and some other network stack will be used. In your case, the OB1 PML is likely selected, and the TCP BTL is used (which does not use Libfabric at all). 

I.E., the root of your problem is what others have stated on this thread - your Libfabric provider is not matching what the OFI MTL is asking for, and things go downhill from there.  You can use mpirun CLI options (noted elsewhere on this thread) to tell the OFI MTL to use RXD and/or RXM and your provider behind it. Then the OFI MTL will enable itself, and the CM PML will enable itself, and then you’re off to the races. 

Now you know what / why. :)

Sent from my phone. No type good. 

> On Nov 13, 2019, at 6:17 PM, Don Fry <DFry at lightfleet.com> wrote:
> I am still trying to crawl, not run a marathon yet.
> Don
> ________________________________________
> From: Hefty, Sean <sean.hefty at intel.com>
> Sent: Wednesday, November 13, 2019 3:15 PM
> To: Don Fry; James Swaro; Barrett, Brian; Byrne, John (Labs); ofiwg at lists.openfabrics.org
> Subject: RE: [ofiwg] noob questions
>> It will need ofi_rxd and/or ofi_rxm since it supports both DGRAM and MSG.
> RxM will need RMA from the msg ep for MPI.
> RxD should be okay with just send and receive.  So I would try to get OMPI running with that, but I wouldn’t expect great performance.
> - Sean
> _______________________________________________
> ofiwg mailing list
> ofiwg at lists.openfabrics.org
> https://lists.openfabrics.org/mailman/listinfo/ofiwg

More information about the ofiwg mailing list