On 18/04/2011 21:23, J Coles wrote:
> This week is the meeting slot usually set aside for the more operations
> team focussed discussions; but this week we want to spend half the meeting
> discussing the GridPP metrics that will be used for the Tier-2 accounting,
> so all site admins are encouraged to attend.
Right, well, we've had a look at
http://pprc.qmul.ac.uk/~lloyd/gridpp/metrics.html at Liverpool. The HS06 Atlas
figure seems OK for Liverpool - most of our nodes are new E5620s, we have a
minority of L5420s, and nothing older than that. Our average HS06 per core is
14.59, so 14.8 seems within a reasonable margin of error given some variation
depending on which nodes jobs land on.
The HS06 Prod figure does seem low. This would appear to be down to our Prod
(s/evt) figure being 105.0 when you would expect it to be significantly lower
given our faster CPUs. So why is it 105.0? I don't know. Where does the figure
come from? What should we be looking into to investigate that?
Then there are some things here which could be factors. We use scaling on our
L5420s for example, scaling them to the HS06 of the E5620s that make up the
majority of our cluster, i.e. they have a multiplier of 0.58 in PBS. Is that
going to affect any of these figures? Not sure.
On top of that, we occasionally overflow jobs on to our local 'tier 3' cluster
when it's idle and we have jobs waiting. This is very much opportunistic use,
just to deliver some additional CPU to the project rather than letting it go
idle. Consequently this isn't in our publishing. Should it be? This would also
affect the available CPU category of the ATLAS metrics. I imagine it's similar
for sites with access to a shared cluster (this would apply to the Computing
Services cluster behind hepgrid3 at Liverpool). Do they publish the entire
cluster, even though in practice they may well not be able to use all of it, or
do they publish only the share they can guarantee, even though at other times
they may well get more? Do they publish the maximum possible CPU share, or the
minimum?
I've also had a look at John Gordon's APEL figures. They confuse me too. Looking
at the EGI accounting portal myself, for ATLAS, in March, Liverpool had a total
normalised CPU time (HEPSPEC06.Hours) of 2,861,280. Totalling up the figures in
John's email, I get 2,145,960. Steve Lloyd's figures have the same, 209.5M.
Where'd the other 715,320 go? Number of jobs is even more confusing. The
accounting portal gives 109,435 ATLAS jobs for Liverpool. John's figures seem to
say we ran 142,884 production jobs alone. Steve's metrics say 50,993 total ATLAS
jobs in March. Eh? Am I doing something wrong here?
I can't help feeling this is all becoming unnecessarily complicated - and that's
without asking about metrics for the other VOs sites get funding for...
--
Robert Fay [log in to unmask]
System Administrator office: 220
High Energy Physics Division tel (int): 43396
Oliver Lodge Laboratory tel (ext): +44 (0)151 794 3396
University of Liverpool http://www.liv.ac.uk/physics/hep/
|