[From nobody Thu Jul  3 14:26:36 2008
Message-ID: <486D3486.7010401@penguincomputing.com>
Date: Thu, 03 Jul 2008 13:20:22 -0700
From: Joshua Bernstein <jbernstein@penguincomputing.com>
User-Agent: Thunderbird 2.0.0.6 (X11/20071022)
MIME-Version: 1.0
To: Joseph Norris <jnorris@ucmerced.edu>
Subject: Re: [torqueusers] infiniban fortran code optimization
References: <486D31AB.20902@ucmerced.edu>
In-Reply-To: <486D31AB.20902@ucmerced.edu>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit

Joseph,

	Is his code written in MPI? Infiniband of course doesn't add an extra 
performance to code magically. Instead it really only optimizes 
throughput and latency for codes that do a lot of inter-node 
communication, such as common with MPI.

	If the code is already implemented with MPI, then well written code 
would not require any changes to run over Infiniband. Instead the MPI 
implementation used just needs to support IB. Most modern MPI 
implementations such as OpenMPI and MVAPICH are capable of running over 
IB. Then, the user just needs to request use of the IB fabric at runtime.

	If the code isn't implemented with MPI, then a major re-write and 
perhaps even a redesign is in his near future. Porting a code from 
non-parallelized implementation to something like MPI (or even just 
multi-threads) can be incredibly challenging. There are TONS of MPI 
introduction available on the web.

-Joshua Bernstein
Software Engineer
Penguin Computing

Joseph Norris wrote:
> Hello to all,
> 
> Have a user that wants to use our infiniban connected nodes with 
> optimization in his fortran code.  Any one have any tutorials/howtos 
> that I can send to him?
> 
> thanks.
> 

]