[MPICH] reg. heterogeneous cluster

Chadalavada Kalyana Krishna kalyanakrishna at gmail.com
Wed Nov 21 01:04:07 CST 2007


Matt,

Thank you for the prompt reply.


"Yes, if you get SMPD to work on your *NIX platform it should be happy to 
talk with Windows SMPDs."

Is there any reason it would not work? I compiled a copy a few mins ago and
was able to start a daemon and run cpi. Strangely, F90 / F77 programs
(pief90, fpi) are little slow to start and ask for input.

Ok ... mpich2 doesn't support automatic endian conversion? .. Any idea when
this is planned?

Thanks,
Kalyan

-----Original Message-----
From: Matt Chambers [mailto:matthew.chambers at vanderbilt.edu] 
Sent: Tuesday, November 20, 2007 10:04 PM
To: Kalyana Krishna; mpich-discuss at mcs.anl.gov
Subject: Re: [MPICH] reg. heterogeneous cluster

Kalyana Krishna wrote:
> Hi everyone,
>
> I have a couple of questions regarding support of heterogeneous
> clusters with mpich2.
>
> 1) With SMPD process manager, is it possible to launch a single MPI
> application across *NIX and Windows platform?
>   
Yes, if you get SMPD to work on your *NIX platform it should be happy to 
talk with Windows SMPDs.


> 2) Does MPICH2 do automatic conversion between machines of different
> binary formats (big endian and little endian) as long as the size of
> the data types is same on both platforms?
>   
No, if you need heterogeneous support you have to stick with MPICH1 
until this feature is added to MPICH2, and the completion date of this 
feature seems to be perpetually pushed back...probably because it's such 
an uncommon requirement. :)


> 3) If answer to 1 & 2 is yes, should the Infiniband driver stack on
> all machines be from the same vendor? (OpenIB, QLogic (Silverstorm),
> Cisco etc)? Or can Linux have OpenIB stack and windows have vendor
> stack and still be able to participate in a single MPI application?
>   
I'll leave this to someone who knows more (i.e. anything) about Infiniband.

-Matt

> thank you for your time.
>
> regards,
> Kalyan
>
>   





More information about the mpich-discuss mailing list