[gmx-users] Benchmarks vs memory bandwidth

[Mike Sullivan]

   I ran a few tests to determine the importance of memory bandwidth. 
The results are rather
interesting. The test case is a 50ps simulation that was given to me by 
a student. The machine
is a Dual Opteron 248 ( 2Ghz ) based on a Tyan S2885 using 1GB DDR 400 
ECC Reg modules.

 This was to determine if the Tyan 2875 was a viable gromacs platform. 
It has only
4 dimm sockets that are attached to one of the two cpus. The second cpu 
must access the
memory pool via the first. You get only two memory channels as opposed 
to four on the
288x series.

Suprisingly even with a single memory channel the performance is similar 
to the
full bandwidth run.

4 memory channels ( 2 dimm Per cpu     )                    23:06
2 memory channels (2 on a single cpu  )                       23:24
1 memory chanel      ( 1 dimm on single cpu )              23:37

FYI for the last case

              NODE (s)   Real (s)      (%)
      Time:   1417.000   1417.000    100.0
                      23:37
              (Mnbf/s)   (GFlops) (ps/NODE hour) (NODE hour/ns)
Performance:    103.784      3.102    127.029      7.872

This was using the 3.2.1 rpms and run on Suse Enterprise 8.0 for AMD64

-- 
Mike Sullivan                           V.P. Performance Computing
@lliance Technologies,                  Voice: (416) 385-3255 x 228, 
18 Wynford Dr, Suite 407                Fax:   (416) 385-1774
Toronto, ON, Canada, M3C-3S2            Toll Free:1-877-216-3199
http://www.alltec.com