From dks@hep.anl.gov Sun Mar  6 08:00:57 2005
Date: Wed, 23 Feb 2005 17:01:58 -0600 (CST)
From: Donald Sinclair <dks@hep.anl.gov>
To: Andrew Siegel <siegela@mcs.anl.gov>
Subject: Re: update on BG/L

I had given up for the moment. Until we get a new version of the compiler
suite that generates sensible code for using the 2 floating-point units,
I will probably not do much on it. (At present if I use the 440d version,
it generates code that gives garbage results.) At present, the per-CPU
performance is abysmal for my code -- about 200 Mflops. Admittedly, I
haven't even tried to get the compiler to align the arrays on quad-word
boundaries. I object to having to implement this as subroutine calls;
it makes the code non-portable. Why couldn't IBM implement it as 
directives, so that it would be ignored by other compilers? I get better
scaling than with other machines, but it is not clear whether this is
due to better communications, or to the fact that the CPU performance is
poor. I figure if I can double the CPU performance, and the network can
keep up, I have jobs that could be run productively, IF I had production
access. (At present I share the 2 general partitions with everyone who
has no partition of their own.) At present I am out of town and dealing
with the BGL is low on my priority list because of this.

                                      Don Sinclair

On Wed, 23 Feb 2005, Andrew Siegel wrote:

> Hi Don: Would you mind giving me an update on your progress on BG/L?? Are
> you running, have you given up, etc? If the former, do you have any general
> performance observations? If the latter, is there any specific
> assistance/tools that might make you more inclined to use the machine. Are
> there any interesting problems that you ultimately foresee being able to run
> on it? Thanks! -andrew 
> 
>  
> 
>  
> 
>  
> 
>