> -----Original Message-----
> From: Testbed Support for GridPP member institutes [mailto:TB-
> [log in to unmask]] On Behalf Of John Gordon
>
> Ewan, what do you mean by 'not altered at sites using CPU Scaling'?
>
I mean, this should be the actual HEPSPEC 06 score as measured, and
not should be scaled at sites that use CPU scaling.
> I believe APEL will use this value if the site doesn't advertise
> cpuscaling.
>
In YAIM we have:
CE_SI00=2171
CE_CAPABILITY="CPUScalingReferenceSI00=2171 Share=atlas:70 Share=lhcb:20 Share=others:10"
CE_OTHERDESCR="Cores=4,Benchmark=8.68-HEP-SPEC06"
Which then appears in the BDII as:
GlueHostBenchmarkSI00: 2171
GlueCECapability: CPUScalingReferenceSI00=2171
GlueCECapability: Share=atlas:70
GlueCECapability: Share=lhcb:20
GlueCECapability: Share=others:10
GlueHostProcessorOtherDescription: Cores=4,Benchmark=8.68-HEP-SPEC06
As I understand it APEL uses the reported (possibly scaled) CPU times
and the advertised (possibly scaled) SI00 values, but not the HEPSPEC06
number from 'GlueHostProcessorOtherDescription'. On a site with no CPU
scaling (like Oxford) those figure should match, in that the SI00 ones
should be 250x the HS06 ones (and in the Oxford example above, they are).
However, if you look at a site like UKI-LT2-QMUL that does use CPU scaling
you get something like:
GlueHostBenchmarkSI00: 1000
GlueCECapability: CPUScalingReferenceSI00=1000
GlueHostProcessorOtherDescription: Cores=12,Benchmark=8.65-HEP-SPEC06
Where the HS06 value is a realistic non-scaled one, but the SI00 figures
are scaled, which is why they're not simply 250x the HS06 number in the
way that the Oxford ones are. And the HS06 value from
GlueHostProcessorOtherDescription is, I think, actually the same as the
value that that Andrew is looking for.
I'm not absolutely sure about this, and possibly we should as one of the
APEL people, but I think that's the way it works.
Ewan
|