As far as I'm aware, we still have an issue with this at Liverpool.
Our cluster right now is still heterogeneous, with a mix of single core Dell
nodes and newer dual quad-core systems. Those latter ones will soon be separated
as they're upgraded to SL5 and attached to a new CE which is currently being set
up, but right now, it's heterogeneous.
Consequently, we use scaling with the Dell nodes used as a reference and the
dual quad-cores scaled to them.
If we calculate an average HEPSPEC06 benchmark, convert that to SI2K, and
publish that as GlueHostBenchmarkSI00, our accounting will be wrong until APEL
actually starts looking at CPUScalingReferenceSI00, as the average and reference
values are not the same.
As far as I'm aware (looking at the savannah bug I've seen mentioned previously,
https://savannah.cern.ch/bugs/index.php?51176 ) APEL hasn't been modified to do
it yet.
So I guess right now we should publish:
The average HEPSPEC06 value in GlueHostProcessorOtherDescription, Benchmark?
The converted HEPSPEC06 value for the reference Dell nodes in GlueCECapability,
CPUScalingReferenceSI00?
And, until APEL is updated to use the above, the same value in
GlueHostBenchmarkSI00? And then when APEL is updated (if that happens before we
split the cluster), we change this to the converted average HEPSPEC06 value?
Does that sound reasonable?
Rob
Stephen Burke wrote:
> Testbed Support for GridPP member institutes
>> [mailto:[log in to unmask]] On Behalf Of [log in to unmask]
> said:
>> 3) We should be publishing:
>> GlueHostBenchmarkSI00: 2500
>> GlueHostProcessorOtherDescription:
>> Cores=4,Benchmark=10.00-HEP-SPEC06
>> GlueCECapability: CPUScalingReferenceSI00=2500
>>
>> Looking at the old scaling numbers and reference of 1000SI2k
>> and the new
>> scaling numbers I think this will overestimate our APEL
>> numbers by about
>> 25%.
>>
>> Is all that reasonable?
>
> Not entirely ... the thing you publish in Benchmark=10.00-HEP-SPEC06 is
> not supposed to be the reference value but the average, i.e. it should
> be possible to calculate the total installed power by multiplying that
> value by whatever you have in LogicalCPUs (which should be the total
> number of cores). Potentially you could take the average as the
> reference but it seems that you don't.
>
> Stephen
--
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/
|