Will X-PLOR run on R10000?

Furio Ercolessi furio at sissa.it
Thu May 16 10:24:20 EST 1996


schuller at indigo1.biomol.uci.edu (Dave Schuller) wrote:
>>  leech at cs.unc.edu (Jon Leech) message <4nbqqr$84h at watt.cs.unc.edu>
>>    <a href="http://www.sissa.it/furio/Mdbnch/results.txt">MDBNCH results</a>
>> show a 195 MHz R10000 Power Challenge running nearly twice as fast as a 266
>> MHz EV5 Alpha.
>
>I cry foul!
>
>That 195 MHz R10000 is in a Power Challenge; a multi-processor
>model line which can be configured with up to 36 processors.  I can find no
>indication on the referenced web page as to how many CPU's the test machine
>contained.  Moreover, the times reported are CPU times, not wall-clock times.
>on multi-processor machines, even if the main program runs on 1 CPU, support
>tasks such as disk I/O may run on other processors and not be counted in the
>total.  Also, no mention is made of amount of RAM, presence of RAMdisks,
>etc on the tested machines which might influence results.

I am the author of the benchmark, although I am not the one who collected
that figure (he appears to be on vacation today).  I will try to help in
this discussion, deferring to a later time the exact details regarding
this measurement.

First of all, one should access the benchmark data from
http://www.sissa.it/furio/mdbnch.html
where a few explanations on the nature of the benchmark can be found.
I can imagine that seeing only the table could only generate perplexities.

MDBNCH is a scalar program, unable to take advantage of parallelism.
As such, it may be used only as an indicator of the speed of a single
CPU on this kind of applications.   I therefore believe the
number quoted for the R10000 processor to reflect the speed of this
processor on this program, and I would expect a similar figure
for a single processor R10000 machine.   This is what happened in the
past for similar cases.  For instance, R4400 Challenge vs R4400 Indigo.

To answer another comment, MDBNCH is a 1988 production program that
was 'benchmarkized' and therefore frozen in its current state at that time.
The reason for its birth is that we wanted to compare machines under
conditions representing our production environment, rather than using "toy 
benchmarks".  Of course our environment somewhat changed from 1988 to today,
but I believe MDBNCH to be still a realistic MD indicator.
It includes runs with up to 16384 particles, which is small 
by today standards but is large enough to exercise cache systems
significantly.  The total memory requirement is of the order of 16 MB,
which is definitely small by today standards: I do not expect a significant
timing difference between a 64MB and a 512MB workstation with the same
processor.
Moreover, it is a MD for metals using a short-ranged many-body
potential, which is probably somewhat different from the typical MD program
used in this community.

>If you want a similar machine to compare to the Power Challenge,
>try the 8400 server, or even the new Alphaserver 4000 line.  Otherwise,
>the appropriate SGI machine to compare to Digital's Alphastation 600
>(which is also available in a 333 MHz version) might be the Indigo R10000.
>I do not find a listing for that machine in the table.

Yes!  I desperately need data! ;-)   I encourage anybody who has access to 
one of these machines to give it a try.  It's really easy, and I am
willing to help.

furio ercolessi
sissa, trieste, italy
furio at sissa.it
http://www.sissa.it/~furio/




More information about the X-plor mailing list