Open MPI logo

Open MPI Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Development mailing list

Subject: Re: [OMPI devel] MXM question
From: Jeff Squyres (jsquyres_at_[hidden])
Date: 2011-08-03 09:11:34


On Aug 3, 2011, at 1:20 AM, Mike Dubman wrote:

> yes, it is for Infiniband only
>
> If OMPI was compiled with mxm support - it will be always selected. (same selection logic as for other MTL components)

That's not quite right for 2 reasons:

1. The other MTL's test their underlying init functions, and if they fail, disqualify that MTL (similar to MXM). PSM does a nice job, for example, returns OPAL_ERR_NOT_AVAILABLE if no PSM hardware is available. MXM does something similar, but it only checks for /dev/infiniband/uverbs0.

This is not sufficient -- it will cause anyone with iWARP or RoCE devices to fail. Indeed, it doesn't even check if the device is a ConnectX device, so anyone with non-ConnectX HCAs may also fail.

Verbs unfortunately fronts multiple different kinds of network devices, so you need to be more careful before unilaterally assuming that the entire world only runs ConnectX devices.

> It should be explicitly disabled to fall back into PML.

I disagree. There's at least 3 cases that I can think of:

1. There's only ConnectX OFED devices in the node.
2. There's only non-ConnectX OFED devices in the node.
3. There's a mix of ConnectX and non-ConnectX OFED devices in the node.

In case 1, MXM should probably assume that it should be used (pending any other tradeoffs vs. openib/ob1...?).

In case 2, MXM should clearly disqualify itself.

In case 3, it's unclear what should happen, IMHO. I have several nodes like this myself, for example. Print a warning message, or no? Allow MXM to continue, or disqualify itself? Error out, or no? ...etc. I'm not sure what the right answer is here.

> On Tue, Aug 2, 2011 at 10:29 PM, Jeff Squyres <jsquyres_at_[hidden]> wrote:
> Is MXM for InfiniBand only?
>
> E.g., what happens if I have the MXM MTL component available on a machine with iWARP or RoCE devices?
>
> --
> Jeff Squyres
> jsquyres_at_[hidden]
> For corporate legal information go to:
> http://www.cisco.com/web/about/doing_business/legal/cri/
>
>
> _______________________________________________
> devel mailing list
> devel_at_[hidden]
> http://www.open-mpi.org/mailman/listinfo.cgi/devel
>
> _______________________________________________
> devel mailing list
> devel_at_[hidden]
> http://www.open-mpi.org/mailman/listinfo.cgi/devel

-- 
Jeff Squyres
jsquyres_at_[hidden]
For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/