Newsgroups: sci.physics.plasma
From johncobb@uts.cc.utexas.edu Wed Jan 4 18:21:55 1995
From: johncobb@uts.cc.utexas.edu (John W. Cobb)
Organization: Oak Ridge National Lab / U. Texas at Austin
Subject: Re: speed of cray vs unix work stations

In article <3ef27h$jj8@mojo.eng.umd.edu>,
<brooker@toka.ireq-ccfm.hydro.qc.ca> wrote:
>At our Fusion lab here, we have the TSC/LSC lower hybrid ray tracing code that
>apparently takes a lot of time on a CRAY so that at most only 3 simulations
>can be done per month. I am thinking of asking that the code be ported to a
>UNIX work station. The million dollar question is how much slower is a powerful
>UNIX work station than a CRAY? A factor of 10? 100? 1000?
>
>Thanks- Peter Brooker
> Centre canadien de fusion magnetique

Of course any answer depends on your particular code, etc. But in my
experience, the correct number is 10 or less. In terms of $/flop, the
smaller the machine, the cheaper. Of course their is a limit on size
below which you cannot run your code, or below which the speed becomes
excessively slow.

As an example, let me speak to some of our experience here at Oak Ridge. I
have several colleagues who, it seems, have been very similar work to yours
concerning RF issues on fusion devices. They used to use Crays exclusively.
Now they are all almost using local RS6k's. The increase in required CPU
time is largish (on the order of 3 to 10) but ironically, the wall clock
time is usually less. The crays we usually work on are the NERSC crays
at Livermore and they have a pretty heavy load. NQS jobs can take days
to queue, especially near APS time. they seem very happy with the arrangement,
especially since our NERSC allocations have been cut about 16%.

In my own experience with a fluid initial value simulation for tilt
stability studies on FRC's [Phys. Fl. B. 5:3227 (1993)] I saw a factor of 4
CPU time difference between a Cray 1s and a low end rs6k (320?) at the
University of Texas at Austin. Your mileage may vary.

Another advantage however, is that the development time can be reduced
considerably. The edit/compile.debug cycle is for me much easier and
faster on workstations. This is true for AIX, but is especially true for
some other vendors where the unoptimized compile-times can even be a
factor of 5 faster. I have friends who can compile on workstations in less
CPU time with a given code then they can on their Crays!!!

If you are thinking of porting a code, my advice is if your problem is small
enough to fit in your memory, do it. If you don't have enough memory, go
buy some more and do it anyway. You can get so much memory loaded on
workstations now that there are only a few problems that are too large
for workstations but not too large for (non-MPP) Crays (or other Vector
Supers).

IMHO, I would suggest using supers only when the job is too large to handle
locally, either because of the size of data files or the amount or crunching.
If your job fits both places, you can get it done faster and cheaper locally,
almost always.

Of course sometimes it is not an apples to apples comparison when you
have to pay for your local cycles out of pocket but supercomputer
access is free or funny money, but that is another topic ....

CYA Note: I do not speak for ORNL, the fusion energy division, the university
of Texas or even myself in any official capacity. These are just musings of
a fellow rambler lost on Usenet.

-John W. Cobb 16% of all Perot voters believe that if Dolphins
are so smart, they should be able to get out of
those nets. --Michael Moore, TV Nation