Hi Daniel, Daniel Persson <daniel@gridcore.se> writes:
Hi list,
I have a really strange problem, i have two different setups, one with Arima motherboards and one with Tyan motherboards. Both are running SLES8 and kernel k_smp-2.4.19-249 with all YAST2 updates. The code im running is Fluent 6.0.18 in parallell mode.(mpich) (CFD code)
The problem is, that the load is very different when running in parallell, on the Arima based cluster, if running on 4CPUS (2dual machines) all cpus are occupied at 80-99% (wich should be considerd normal). However, running the exact same code and test case on the Tyan based machines(same kernel, network etc) then 2CPU is working hard and the other 2 are "watching" and this is one CPU per machine, ie one cpu in the dual machine is working, and CPU2 are idling. The result is that the same code and case are about 10times slower on the Tyan machines. (wich have the latest BIOS)
The only(important?) difference in the setup is that the Tyan based test are connected to a RedHat 8.0 boxwith Dual Xeons. i.e auth and nfs are
Replace it with a machine running SuSE ;-). Seriously, this shouldn't be a problem.
from this machine. But all computations are done on the Opteron machines.
So, basically, my question are, could i have done some serious config mistake, or is there any problem with the kernel/Tyan bios wich casues one CPU to take alot more load than the other. Can i config the CPU routing somehow ?
You can bind processes to specific cpus via the /proc/interface or the numactl program. Are you using 2 2p boards in each case? Or a 4p one?
On the tyan machine, it also takes long time to just read the casefile.(ide disks, DMA and 32bitaccess enabled, same as Arima setup)
Any ideas anyone ?
Andreas -- Andreas Jaeger, aj@suse.de, http://www.suse.de/~aj SuSE Linux AG, Deutschherrnstr. 15-19, 90429 Nürnberg, Germany GPG fingerprint = 93A3 365E CE47 B889 DF7F FED1 389A 563C C272 A126