Re: : PROOF

Hi Isidro,

   don't use NFS for PROOF file distribution!

iglez_at_mail.cern.ch wrote:
> Hi Gerardo,
> I have another few questions related to PROOF:
>
> 1) I wonder if PROOF is clever enough to process the events where the data
> is local. The place where the data is seems to be quite important. I
> have tried two different configurations:
>
> a) All data in the local disk of the master exported through NFS to all
> slaves (the same path name in all of them). It slows down when I add
> slaves!!!
>

The more slaves the more contention you will get on the single or few NFS servers that have to serve all the files, hence the slower PROOF goes with more slaves.

> b) The data is distributed through the slaves and all of them export
> their disk through NFS to all the rest (including the master). The time
> consumed now is reduced. But still, adding more slaves not always
> improves the performance.
>

When using NFS, PROOF does not know the file location and cannot use any optimizations based on the file location. In the above case there will still be NFS servers serving files for more than one slave.

> 2) Mi conclusion is that the NFS is killing the improvement. Is there a
> way to tell root (for example when building the TDSet) which data file
> should be analysed in which machine?
>

You should only access files via the rootd protocol on PROOF clusters. Using rootd files are specified in TDSets like:

    root://host1/data/file1.root
    root://host2/data/file2.root
    etc.

in which case PROOF uses the host info to make sure as much as possible local data is accessed by the slaves.

Cheers, Fons.

> Cheers,
>
> Isidro
>
>
> --
> +---------------------------+--------------------------------+
> | Isidro González Caballero | mailto:Isidro.Gonzalez_at_cern.ch |
> | IFCA (U. Cantabria/CSIC) | http://home.cern.ch/iglez/ |
> | Avda. Los Castros s/n | -o- |
> | Santander (España) | Tel: +34 942201343 / 699413841 |
> +---------------------------+--------------------------------+

-- 
Org:    CERN, European Laboratory for Particle Physics.
Mail:   1211 Geneve 23, Switzerland
E-Mail: Fons.Rademakers_at_cern.ch              Phone: +41 22 7679248
WWW:    http://www.rademakers.org/fons/      Fax:   +41 22 7679480