[Beowulf] Thought that this might be of interest

Richard Walsh rbw at ahpcrc.org
Mon Nov 6 12:19:48 PST 2006

Joe Landman wrote:
> Richard Walsh wrote:
>> Joe Landman wrote:
>>> We are working on a machine with 2 Opteron 2218s.  For laughs, I ran 
>>> streams on it.  Here are 1,2,3,4 processor data.
>>> 1 processor:
>>> Copy:       5713.9944      0.0591      0.0560      0.0600
>>> Scale:      5713.9822      0.0587      0.0560      0.0600
>>> Add:        5454.2389      0.0911      0.0880      0.0920
>>> Triad:      5454.1576      0.0916      0.0880      0.0920
>>     Mmmm ... these actually seem low considering the bus bandwidth is 
>> supposed to be 10.67 GB/sec
> Heh.... don't shoot the measurer ... :)
    Never measurers are honored guests at my table ... ;-)
> I agree BTW that I expected this to be about 7++ GB/s per socket.  
> This is a pathscale built binary.  Will try with PGI and Intel as well.
     In the presentation I directed folks to there are socket F stream 
numbers for both 2x2 and 4x2 cases (4 and
     8 threads).  These are AMDs numbers, but they report:

     2x2  4 threads  Intel 5160  3.0 GHz             at   5865 for the 
triad (27% of  maximum)
     2x2  4 threads  Opteron 2220 SE 2.8 GHz  at 13296 for the triad (62 
% of maximum)
     4x2  8 threads  Opteron 2220 SE 2.8 GHz  at 18271 for the triad (42 
% of maximum)

     These numbers are both good and bad.  Good in the sense that socket 
F shows fairly
     good percentages of designed bandwidth.  Bad in the sense that one 
might expect--less
     and less of it seems usable as the core/thread count go up.   A 2x4 
system I would expect
     to do better than an 4x2 system on a percentage basis, but the 
channels are a bit slower
     on Barcelona (back at socket 940 6.4 GBs levels), so the absolute 
number may be the
     same or worse.



Richard B. Walsh

"The world is given to me only once, not one existing and one
 perceived. The subject and object are but one."

Erwin Schroedinger

Project Manager
Network Computing Services, Inc.
Army High Performance Computing Research Center (AHPCRC)
rbw at ahpcrc.org  |  612.337.3467

This message (including any attachments) may contain proprietary or
privileged information, the use and disclosure of which is legally
restricted.  If you have received this message in error please notify
the sender by reply message, do not otherwise distribute it, and delete
this message, with all of its contents, from your files.

More information about the Beowulf mailing list