Hi Stephen
> LHC Computer Grid - Rollout
>
>> [mailto:[log in to unmask]] On Behalf Of Gonçalo Borges said:
>> The majority of the nodes (60%) has a 6.8625 HEP06 value per
>> core which
>> converted to SI00K gives 1716 and this is the value published as
>> GlueCECapability: CPUScalingReferenceSI00=1716.
>>
> One comment, I don't think the version of APEL which reads that attribute is available yet:
>
> https://savannah.cern.ch/bugs/index.php?51176
>
>
Just to clarify, what I really meant was:
goncalo@lnsys15 ~]$ ldapsearch -x -H ldap://sbdii01.ncg.ingrid.pt:2170
-b mds-vo-name=NCG-INGRID-PT,o=grid | grep GlueHostBenchmarkSI00:
GlueHostBenchmarkSI00: 1716
Thanks for the explanations...
Cheers
Goncalo
>> You can see that the fact that we are not doing nothing to
>> the wallclock
>> time for those faster machines (in the previous document
>> nothing is said
>> about that) it gives origin to efficiency values higher than
>> 100%. Also,
>> I do not think that something SHOULD be done to the wallclocktime
>> because an application can be very I/O intensive, and in that
>> case, it
>> would take approximately the same time in a fast or slow CPU.
>>
> Personally I agree, but I think the general practice is to scale the walltime too. Anyway the walltime limits should be quite a long way above the CPU time limits, the purpose is mainly to catch jobs which have got stuck and are consuming no CPU.
>
>
>> The question is: Are these efficiency values reasonable to be
>> presented?
>>
> Well, they are genuine, you are calculating your CPU power in a unit related to your slower CPUs, so the faster ones can indeed deliver more than 100% of that, i.e. more than one hour of power units per wallclock hour. On the other hand it may confuse people and you'll probably have to keep explaining it ...
>
> Stephen
>
|