Re: Octave: Question on running parallel package on linux HPC cluster

[Tung Nguyen]

Thanks a lot for your input Carlo and Olaf! qsub will give me the computed nodes with their associated CPUs name. But it's seemed that they all have the same names as in my first post. I will have to dig further into this to find out how I can access individual CPUs and use them together with "parallel" package

Thanks again for helping!

-- Tung

On Wed, Sep 23, 2015 at 12:00 PM, Olaf Till <address@hidden> wrote:

On Tue, Sep 22, 2015 at 05:57:17PM +0000, Carlo De Falco wrote:
>
> On 14 Sep 2015, at 16:05, Olaf Till <address@hidden> wrote:
>
> > I'm not familiar with qsub (gridengine?), but it has a different
> > concept and you don't need the parallel package for it.
>
> Actually qsub is part of the Torque resource manager:
>
> http://www.adaptivecomputing.com/products/open-source/torque/
>
> usually on clusters using a job scheduler / resource manager you
> need qsub to launch any job, be it serial of parallel.
>
> So if you want to have multiple cores to run a parallel Octave
> script using either the parallel or mpi package you will need
> to write any instructions needed to run your script in a submission
> file then launch it through qsub.
>
> HTH,
> c.

So I should correct myself: The parallel package might be usable in
this case, but I can't give much detailed advice since I don't know
this qsub system. So the OP probably has to work out the details
himself, using the documentation in 'parallel'. parcellfun/pararrayfun
might be usable, or, if the nodes should have separate ip-addresses (I
don't know this), netcellfun/netarrayfun. But in the latter case
pserver would have to be started at each node first; if the nodes are
not accessible by anyone else as long as the pservers run, TLS might
not be necessary, but the OP has to assess this.

Olaf

--
public key id EAFE0591, e.g. on x-hkp://pool.sks-keyservers.net