[mvapich-discuss] Obtaining older releases

[Lukasz Zalewski]

On 26/01/14 23:52, Jeff Hammond wrote:
>> Hi DK,
>> We are trying to use an alternative mpi build for our Matlab installation on
>> ib cluster and we are limited w.r.t the versions that we can use
>> (http://www.mathworks.co.uk/help/mdce/use-different-mpi-builds-on-unix-systems.html).
>>
>> It seems that mvapich2 1.8, which is based on mpich2 1.4.1p1, is the latest
>> that we can use for this endeavour, as 1.9 and 2.0b releases have been based
>> on mpich 3.x.x.

Hi Jeff,
The issue is related to mpich, rather than mvapich. For completeness 
sake I shall describe my issue here and then move to the suggested mpich 
list for further discussion.

>
> What is the issue that compels you to use an MPICH2 1.4.1p1
> derivative?  It's not clear why you are stuck.  You'd be stuck with
> this version if you were trying to use Windows, but that's clearly not
> the case given the link you shared.

Matlab by default ships with mpich2 1.4.1p1. As it does not have hydra 
pm support we decided to look for an alternative. We tried latest mpich 
(and mvapich) but noticed that it causes a crash in the matlab workers. 
However custom built mpich2 1.4.1p1 and mpich2 1.5 worked out of the box.

I presume that the problem is caused by the ABI incompatibilities 
between mpich2 and mpich 3.x?

>
> The right approach to your problem is almost certainly to address
> whatever feature of more recent MPICH and derivatives that is blocking
> you and not to use out-of-date, unsupported code.
>
> Please feel free to post your MPICH 3.x issues to discuss at mpich.org.
> If your issues are not specifically IB-related, you should try to
> debug them upstream.  It should not be hard for the OSU folks to merge
> MPICH-generic patches downstream.

I'm quite happy moving this discussion and providing more information if 
necessary.

>
> I notice on the Mathworks page that they document the use of gforker
> and mpd.  The mpd process manager has been deprecated for a long time
> and is no longer supported.  There is very little reason not to use
> Hydra.  If assuming a process manager other than Hydra is what blocks
> you, you should fix that issue.

We do want to use hydra, hence our need for custom built mpich (and 
mvapich for ib support).

Regards

L