On Thu, Jun 25, 2009 at 09:59, A.J.Martin<[log in to unmask]> wrote:
>
>>>
>>> PS We seem to have overtaken Glasgow :-).
>>
>> Aggghh, damn you! Of course, we ran out of jobs so we need more work...
>>
>> What's more interesting, of course, is the analysis throughput we each
>> get. I want to know if you get substantially better throughput than us
>> with your whizzy lustre vs. our humble rfcp to local disk.
>>
>
> According to our accounting the jobs that ran yesterday had an cpu/wall
> time efficiency of ~64% . Assuming I'm reading out the switch correctly the
> traffic from the storage was peaking at ~700 MB/s with ~500 jobs running. We
> believe we can achieve ~5 times that, so it will be interested
> in seeing what happens when the cluster is fully loaded.
Have a look at:
http://atlas-ganga-storage.cern.ch/test_479/
Glasgow and QMUL both got 60% CPU efficiency and a 10Hz event rate.
This is a very simple i/o intensive AOD analysis. You're right about
wanting to scale up further and also test sparse TAG analysis.
We had some failures to stage-in data which we need to dig into and
find out if we had a flaky disk server last night.
cheers
Graeme
--
Dr Graeme Stewart http://www.physics.gla.ac.uk/~graeme/
Department of Physics and Astronomy, University of Glasgow, Scotland
DEATH TO MEETINGS!
|