Open MPI logo

Open MPI Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Development mailing list

Subject: Re: [OMPI devel] Heterogeneous OpenFabrics hardware
From: Jeff Squyres (jsquyres_at_[hidden])
Date: 2009-01-27 15:24:16


There was discussion about this issue on the call today. Several
vendors expressed willingness to "fix" the problem and make OMPI
interop with different HCAs and RNICs in a single job run: Sun, IBM,
Mellanox.

So the question is -- what exactly do you want to do to fix this?
Assumedly this kind of auto-detection will only occur when we have a
modex -- further assuming that each port (module) can pass around its
vendor and part ID. Then upon start_connect(), you can look up your
peer's vendor and part ID and look up its values in the INI file and
see if there's a clash. That's a very general idea; I haven't re-
looked at the code recently to see if that would actually work.

I posted one idea here:

     http://www.open-mpi.org/community/lists/users/2009/01/7861.php

But I'm not entirely convinced that's the Right way to go.

On Jan 27, 2009, at 8:15 AM, Jeff Squyres wrote:

> On Jan 26, 2009, at 4:46 PM, Jeff Squyres wrote:
>
>> Note that I did not say that. I specifically stated that OMPI
>> failed and it is due to the fact that we are customizing for the
>> individual hardware devices. To be clear: this is an OMPI issue.
>> I'm asking (at the request of the IWG) if anyone cares about fixing
>> it.
>
>
> I should clarify something in this discussion: Open MPI is *capable*
> of running in heterogeneous OpenFabrics hardware (assuming IB <-->
> IB and iWARP <--> iWARP, of course -- not IB <--> iWARP) as long as
> it is configured to use the same verbs/hardware configuration on all
> the hardware. Depending on the hardware, Open MPI may not be
> configured to run this way by default because it may choose to
> customize differently for different HCAs/RNICs.
>
> However, if one manually configures Open MPI to use the same verbs/
> hardware configuration values across all the HCAs/RNICs in your
> cluster, Open MPI will probably work fine. If Open MPI doesn't work
> in this kind of configuration, it may indicate some kind of vendor
> HCA/RNIC incompatibility.
>
> Case in point: I regression test "limited heterogeneous" scenarios
> on my MPI testing cluster at Cisco every night. Specifically, I
> have a variety of different models of Mellanox HCAs and they all
> interoperate just fine across 2 air-gapped IB subnets. I don't know
> if anyone has tested with wildly different HCAs/RNICs using some
> lowest-common denominator verbs/hardware configuration values (i.e.,
> some set of values that is supported by all HCAs/RNICs) to see if
> OMPI works. I don't immediately see why that wouldn't work, but I
> haven't tested it myself.
>
> Out of the box, however, Open MPI is not necessarily configured to
> have the same verbs/hardware configuration for each device. That is
> what may fail by default.
>
> --
> Jeff Squyres
> Cisco Systems
>
> _______________________________________________
> devel mailing list
> devel_at_[hidden]
> http://www.open-mpi.org/mailman/listinfo.cgi/devel

-- 
Jeff Squyres
Cisco Systems