[Beowulf] Dual Core: real live data

[Steve Cousins]

I have been testing our code out on a Dual Opteron 275 system and it was
very helpful.  Based on my findings I'm planning on getting two separate
dual 248 (or higher) systems instead of one Dual 275 system. 

=======================================================

Single program running:	
		4 Cores		2 Cores		1 Core
Time (seconds)	509		739		1247

=======================================================

Two Dual-core jobs running at the same time:
		Job A		Job B
Time (seconds)	878		882	

=======================================================

Four Single-core jobs running at the same time:
		Job A	Job B	Job C	Job D
Time (seconds)	1585	1522	1580	1545

=======================================================

What I was primarily interested in was comparing the time it took two
2-core jobs to run vs. the time for one 2-core job to run. There appears
to be a hit of about 20% from 739 seconds to 880 seconds.  This is why I
am leaning towards the two separate machines.  

Sure it would be nice to run a single 4 core job once in a while but the
primary use will be to have two people working on their own models.  

Just because it was there, I did:

=======================================================
Four Single-core jobs running at the same time:
		Job A	Job B	Job C
Time (seconds)	1593	1251	1625
=======================================================

I believe Jobs A and C shared a CPU.  Why they took longer than the
Four single-core jobs I don't know.  This fairly clearly shows the impact
of having two jobs running on the same core, at least for this
application.

Thanks very much to Jeff Nguyen at ASL Inc. (www.aslab.com) for providing
the machine to test out.  I had never heard of ASL or Jeff before this. I
have no affiliation with them whatsoever.  Jeff asked me if I wanted to
test it out and that was it.  No questions asked and no sales junk pushed
on me. Just for curiousity I asked how much one of the dual 275 systems
would be in order to compare with the price of putting one together
myself.  It seemed very reasonable.

When I got done, I emailed Jeff and told him my results.  He then offered
to swap out the CPU's with two 248's so I could compare with identical
hardware (except the CPU's of course):

=======================================================
Single program running:	
		2 Cores		1 Core
Time (seconds)	750		1378
=======================================================
Two programs running:
		1 CPU (A)	1 CPU (B)
Time (seconds)	1328		1344
=======================================================

For 2 CPU's, the 248's seem to work about the same as two cores on the
Dual 275 if you run with one core from each of the 275's.  For single CPU
jobs, it seems that the 275's are about 7 to 10 percent faster than the
248.  I don't know how to explain this.  Could be that some cron job
kicked in in the middle of the run.  The only thing I see in the logs is
sar every 10 minutes.

These results are hardly scientific, however they are useful to me and I
hope to others.

Steve
______________________________________________________________________
 Steve Cousins, Ocean Modeling Group    Email: cousins at umit.maine.edu
 Marine Sciences, 208 Libby Hall        http://rocky.umeoce.maine.edu
 Univ. of Maine, Orono, ME 04469        Phone: (207) 581-4302