ATHLON vs XEON: number crunching

Steffen Persvold sp at scali.com
Fri Jun 21 10:16:33 PDT 2002


On Fri, 21 Jun 2002, Thomas Guignon wrote:

> Le Jeudi 20 Juin 2002 15:06, vous avez écrit :
> > Hi Steffen,
> >
> > Looking at your numbers I can make a conclusion that loading 2 processors
> > on Xeon simply slows down the stream as for Athlon it does not and they
> > are quite close to each other in case of full load of the node. Are you
> > saying that the real test would be to run TWO IDENTICAL codes on the same
> > node to load both processors?
> >
> > Ivan
> 
> Hello
> I would be interesting to know which compiler was used for this stream 
> benchmark? probably is the Intel compiler.
> It's possible to improve the AMD numbers if prefetch is not in use but it 
> will change the scalability conclusion of  Ivan Oleynik
> The fact that dual Xeon perf decrease is quite strange (no speedup would be 
> more normal) and there is some investigation to do:
> - Does the  2 processors test was obtained with openMP (I guess it is) ?

IIRC it was the 3.2-4 PGI compiler with SSE and OpenMP switches. 
However, if I run with only 1 OpenMP thread, but two instances (as you 
would do without the OpenMP capability) I get the following

Instance 1:

Function     Rate (MB/s)  Avg time   Min time  Max time
Copy:        583.8777      0.0615      0.0548      0.0637
Scale:       672.0295      0.0539      0.0476      0.0561
Add:         656.5718      0.0784      0.0731      0.0818
Triad:       670.0255      0.0794      0.0716      0.0834

Instance 2:

Function     Rate (MB/s)  Avg time   Min time  Max time
Copy:        544.0237      0.0613      0.0588      0.0642
Scale:       613.8390      0.0550      0.0521      0.0580
Add:         653.0164      0.0798      0.0735      0.0854
Triad:       645.2300      0.0782      0.0744      0.0821


If you manually add the number you get roughly the same as with OpenMP.

Regards,
Steffen
 > 
> > ------------------------------------------------------------------------
> >
> > On Thu, 20 Jun 2002, Steffen Persvold wrote:
> > > Well, the E7500 chipset has two DDR _channels_ while the AMD762 system
> > > controller has one DDR _channel_. One DDR channel is 64bit + ECC and is
> > > normally running at 133MHz which gives us a theoretical peak performance
> > > of 2.1 GBytes/sec per channel. However, the E7500 chipset is only able to
> > > run the memory bus at 100MHz and with the interleaving of the two
> > > channels that gives us a theoretical peak performance of 3.2 GBytes/sec.
> > >
> > > I believe the term "banks" is related to the layout of the DIMMs and has
> > > nothing to do with the architecture of the system controller.
> > >
> > > Here are some stream numbers. The Athlon platform is Tyan S2466 equipped
> > > with two Athlon MP 1900+ (1.6 GHz), and the Xeon platform is SuperMicro
> > > P4DPR-6GM+ equipped with two Xeon 1.8 GHz (Prestonia).
> > >
> > > Athlon, 1 process :
> > >
> > > Function     Rate (MB/s)  Avg time   Min time  Max time
> > > Copy:        877.5785      0.0365      0.0365      0.0367
> > > Scale:       801.7857      0.0400      0.0399      0.0402
> > > Add:         894.9042      0.0537      0.0536      0.0539
> > > Triad:       887.2120      0.0542      0.0541      0.0544
> > >
> > > Athlon, 2 processes :
> > >
> > > Function     Rate (MB/s)  Avg time   Min time  Max time
> > > Copy:       1385.1016      0.0236      0.0231      0.0253
> > > Scale:      1345.9527      0.0242      0.0238      0.0250
> > > Add:        1168.2247      0.0419      0.0411      0.0425
> > > Triad:      1210.0117      0.0405      0.0397      0.0473
> > >
> > >
> > > Xeon, 1 process :
> > >
> > > Function     Rate (MB/s)  Avg time   Min time  Max time
> > > Copy:       1898.8954      0.0170      0.0169      0.0181
> > > Scale:      1813.9614      0.0179      0.0176      0.0186
> > > Add:        1839.8592      0.0265      0.0261      0.0274
> > > Triad:      1826.2738      0.0266      0.0263      0.0274
> > >
> > > Xeon, 2 processes :
> > >
> > > Function     Rate (MB/s)  Avg time   Min time  Max time
> > > Copy:       1160.5962      0.0297      0.0276      0.0311
> > > Scale:      1261.8243      0.0269      0.0254      0.0275
> > > Add:        1237.7180      0.0395      0.0388      0.0407
> > > Triad:      1231.9053      0.0398      0.0390      0.0405
> >
> > ------------------------------------------------------------------------
> > Ivan I. Oleynik                       E-mail : oleynik at chuma.cas.usf.edu
> > Department of Physics
> > University of South Florida
> > 4202 East Fowler Avenue                  Tel : (813) 974-8186
> > Tampa, Florida 33620-5700                Fax : (813) 974-5813
> > ------------------------------------------------------------------------
> >
> > _______________________________________________
> > Beowulf mailing list, Beowulf at beowulf.org
> > To change your subscription (digest mode or unsubscribe) visit
> > http://www.beowulf.org/mailman/listinfo/beowulf
> 
> 




More information about the Beowulf mailing list