[gmx-users] (no) superlinear Scaling on Opterons

[Jens Krueger]

Hello,

there have been multiple reports on the list of superlinear scaling on
AMD Opterons of the 200 series.

We have recently purchased a small cluster with the following
configuration for each compute node:
- 2x Opteron 2218 (4 cores per node)
- 2 GB RAM
- Supermicro H8DM8 mainboard
- Mellanox PCI-E x8 Infiniband (running with OFED-1.2 and mvapich2)
- Rockscluster 4.3
- Gromacs-3.3.1


For the DPPC benchmark we achieve the following numbers:

cores ns/day speedup

01 0.289 1.00
02 0.487 1.69
04 0.768 2.66
08 1.330 4.60
12 1.776 6.15
16 2.137 7.39
20 2.550 8.82


Compared to the benchmark section
(http://www.gromacs.org/content/view/25/38/) and other published data
(e.g. http://biowulf.nih.gov/apps/gromacs/bench-3.3.1.html) the
performance drops to soon when using more cpu's. In fact we are far away
from superlinear scaling and the performance on 20 cores is only half as
good as on the biowulf cluster.
We systematically checked the hardware and tried different compiler and
MPI, all giving no or only minor improvement.

Has anybody hands on an Opteron cluster with cpu's from the 2000 series?
It would be of immense value for us to get reference data from a similar
machine.


Thanks,

Jens


-- 
Dr. Jens Krueger

Institute of Biophotonics
School of Biomedical Science and Engineering
National Yang-Ming University
155, Li-Non St., Sec. 2
Taipei, 112, Taiwan (R.O.C.)

mobile: 0921452940

http://www.ym.edu.tw/~wfischer

-------------- next part --------------
A non-text attachment was scrubbed...
Name: jk.vcf
Type: text/x-vcard
Size: 146 bytes
Desc: not available
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-users/attachments/20071004/a9303d99/attachment.vcf>