On Thu, Jun 24, 2004 at 02:10:30PM -0700, Phil Ehrens wrote:
>Luke Palmer wrote:
>> to have four logical processors each due to hyperthreading. Would there
>> be any advantage to starting four lam processes per node, rather than
>> two, or is this just silly?
>You should do a performance test, but I would say it's
>probably a loser.
For us it seems to be a win for single node OpenMP, but a loss for MPI.
This is due to Linux's imperfect scheduling, so it should get better as
the scheduler gets more HT aware.
There are a couple of alternate 2.6 schedulers at the moment - perhaps
you could try them out and see how they handle HT?
The case that kills you is when there are 2 threads on a SMP node
(eg. 2 physical cpus == 4 logical) and if the scheduler puts both
the threads onto the same physical cpu.
The 2 running threads (and 2 sleeping threads) case actually occurs
fairly often as MPI load balancing is rarely perfect.
cheers,
robin
|
