Hello,
Our cluster nodes has 2 SMP CPU and with hyperthreading giving 4 CPU. But
linux scheduler always switch the CPU of LAM process, causing decreasing of
performance (HPL result cut half).
So we use "taskset" utility to start LAM process. This is done by using
schema:
n0 -np 1 taskset 1 ./xhpl
n0 -np 1 taskset 8 ./xhpl
n1 -np 1 taskset 1 ./xhpl
n1 -np 1 taskset 8 ./xhpl
...
This works, but have some inconvenience. To change the CPU used by the job,
we need to change the schema file too. This goes agaist that when we use
PBS TM startup, everything is specified through parameters of qsub (no need
to change mpiexec line in job script).
Is there a smarter way to use taskset in LAM startup? Or is there an option
to set affinity in LAM internally?
Thanks.
Ricky Tang
Cluster Technology Ltd
|
