On 2 Nov 2007, at 10:57, Burke, S (Stephen) wrote:
>
>> Since _LCG_ were never interested in SMP processing, there wasn't a
>> penalty in making the site look like a larger number of
>> smaller single-CPU boxes.
>
> There may however be some interest from other VOs in knowing the real
> situation ... I can see the logic in this kind of thing but it also
> inclines me to despair, because the implication is that it's pointless
> to try to fix things which are broken, the best we can do is break
> something else to compensate!
I think it will become important to sort this properly, but at the
moment the big block is the lcg-RB and CE which deal with MPI jobs in
an appalling (and hard coded) way. Stephen Childs has put in a lot of
effort to work around the most egregious flaws and more or less get
MPI to work on EGEE clusters.
With the glite-WMS now being rolled out, then I think the job
wrappers for MP(I) can be more flexibly, and sensibly, implemented.
MPI might not be of much interest to HEP, but with the core explosion
on chips MP jobs might well be.
Make a fresh start with Glue2? Publishing values per job slot and
values per machine might be sensble.
cheers
g
--
Dr Graeme Stewart - http://wiki.gridpp.ac.uk/wiki/User:Graeme_stewart
ScotGrid - http://www.scotgrid.ac.uk/ http://scotgrid.blogspot.com/
|