[MITgcm-support] MITGCM on dual-dual-core Opteron
Jason Goodman
jgoodman at whoi.edu
Thu Jun 15 17:00:33 EDT 2006
Hey, MITGCM gang!
My group here at WHOI has a 28-node cluster with dual Opteron 248
nodes. We're going to be replacing some of them with dual-core
Opteron 275s (same clock rate, but dual-core), and I've been doing
some benchmarking on a trial node.
MITGCM ocean model build 20050721
3-d convection code (verification/exp5)
domain = 120x120x20
Dual core AMD 275:
Cores Usr Time Rate
------------------------------------
1 133.3 1
2 120.0, 116.5 1.11, 1.14
4 77.4 1.72
Single core AMD 248:
Cores Usr Time Rate
------------------------------------
1 132.2 1
2 87.1 1.52
4 (2 nodes) 55.6 2.38
Now, I don't expect MITGCM to work well with dual-core chips. I've
been told that MITGCM is memory-bandwidth-limited, and so when two
cores on the same chip try to access their local memory at the same
time, there's a traffic jam on the memory bus. Compare with NCAR
CAM, an atmospheric model which is CPU-number-crunching limited:
Dual core AMD 275:
Cores Wallclock time Rate
------------------------------------
2 797 1
4 394 2.0
Single core AMD 248:
Cores Wallclock time Rate
------------------------------------
2 775 1
4 (2 nodes) 410 1.9
I'd like to be able to use the quad-core nodes as dual-core nodes
sometimes, but MITGCM's performance using 2 cores per node is
terrible. The 275 chip has the same clock speed as the 248, so you
would think that each MITGCM process would run on a separate chip,
each with its separate memory, and performance would be equal to the
248. That's exactly what happens with CAM. But with MITGCM on my
system, the processes hop around from CPU to CPU. Sometimes both
processes will be on the same chip, sometimes they'll be on different
chips, the scheduler moves them around like crazy. And of course,
when they're on the same chip, you get a memory bottleneck.
Why does this happen? Why MITGCM and not CAM? Is there anything I
can change in my OS configuration or MITGCM configuration to stop it?
Jason
More information about the MITgcm-support
mailing list