Re: sgi

[Geoff Jacobs <gdjacobs@gmail.com>]

Adam C Powell IV wrote:

>On Wed, 2005-10-19 at 18:35 +0000, Konstantinos P wrote:
>  
>
>>Adam C Powell IV wrote:
>>
>>    
>>
>>>Indeed, PVM can also run a single parallel job across a heterogeneous
>>>cluster.  But I know that mpich can't do that for MPI parallel jobs
>>>(http://bugs.debian.org/31779), and I don't believe lam can either.
>>> 
>>>
>>>      
>>>
>>I have never use MPI in heterogeneous clusters, but what about
>>specifying the architecture
>>using the arch argument of mpiexec?
>>The mpiexec manual page in the LAM MPI distribution says that the user
>>can specify
>>the architecture using this argument.
>>If you use LAM you can check your architecture string with laminfo. For
>>example in my computer
>>this returns i386-pc-linux-gnu. In case of SGI wouldn't this return
>>something like mips-sgi-linux-gnu?
>>If this is true would still be a problem to use MPI in a heterogeneous
>>cluster?
>>    
>>
>
>Sounds like I stand corrected regarding LAM.  But MPICH does not have
>mpiexec, though mpich2 does (not yet a Debian package), so I don't think
>mpich can do this.
>
>Thanks,
>-Adam
>  
>
Actually, with MPICH1 when you use the p4 device, you can write a custom
p4 script which lets you run specific binaries on specific machines. I
used this for asymmetric tasking, but you could also use this for a
multi-arch run if you have binaries compiled for each architecture.

I'm not sure if you'd have problems with endian-ness and whatnot. I
believe MPICH is supposed to handle that transparently, but I've seen
some weird shit before, so I don't really know until someone smacks down
the beef.

-- 
Geoffrey D. Jacobs
MORE CORE AVAILABLE, BUT NONE FOR YOU.