Stephen Burke wrote:
> Testbed Support for GridPP member institutes [mailto:TB-
>> [log in to unmask]] On Behalf Of John Gordon said:
>> The main problem is that APEL accounting which can normalise cpu using
>> a site average doesn't know which are production and which analysis
>> jobs.
>
> Doesn't it store the DN and FQAN the job was run under?
>
>> The site average HEPSPEC06 is already published in the BDII though.
>
> The subcluster average in fact - so if sites create a subcluster per hepspec category it's accurate, but of course not everyone does.
QMUL, for one, does not. Apart from anything else, we don't want jobs
queuing when we have free slots.
Hyperthreading will have a large and difficult to measure impact -
though it wasn't really discussed at GridPP. QMUL and Manchester have
this turned on - and if you look at Manchester's hepspec06 results,
there is a 20% increase in machine hepspec to be gained. Experiments
will almost always gain - sites won't unless the cluster is full.
Assuming that hepspec is representative[1], it would be extremely silly
if the accounting metrics pushed sites to turn this off (and potentially
lose nearly 20% of the UK's installed capacity).
Chris
[1] With analysis jobs waiting on IO, I suspect we'd do even better than
that.
|