[mpich-discuss] MPICH2 CPUs Vs. Cores

[Reuti]

Hi,

Am 15.07.2011 um 23:34 schrieb Peter Willis:

> I have a single system with 4 CPUs. Each CPU has 12 cores.(48 cores total)
> MPICH2 is built and installed from source and there are no errors.
> 
> The documentation is a bit vague regarding CPUs  vs. Cores.
> 
> QUESTIONS:
> 
> 1.)
> What is the best way for a single node system to utilize
> a maximum number of cores?

what do you mean by this in detail? Do your applications scale well with increasing number granted cores?


> 2.)
> Do I need PBS or Hydra to run jobs on a single node?

These are two different things. PBS is a batch scheduler, and it will allow you to serialize your workflow (unless you will use 48 cores all the time for sure). Means you can submit any set of 8, 16 or whatever core jobs, and they will be started when enough cores are available again. This makes more sense than starting Friday evening jobs with 128 cores and oversubscribe the machine over the weekendand make them fight for memory. Instead any combination of them will run at a time ionce they are submitted to the queuing system, so that cores and memory are well used and not oversubscribed.

Biased opinion: I would suggest to use Grid Engine instead, but both will do like Torque or LSF.

If all is well set up, the parallel jobs will just fork, as no task is has to be started on any slave node. Hydra is installed anyway with MPICH2 and won't hurt. 

-- Reuti