Hi,
To have another go at the LogicalCPUs question, maybe it helps to look
at it from the other end. What we want to achieve here is to know the
total CPU power of a site, to compare with the pledges to GridPP and
LCG. That power will be calculated as HEPSPEC (taken from
ProcessorOtherInfo) multiplied by LogicalCPUs. Hence those numbers need
to be set so that their product gives a fair picture for your site. In
general the assumption is that sites run one job per core, so
LogicalCPUs counts cores and HEPSPEC is calculated by running the
benchmark once per core. If the site does something different it may
need to adjust that. So e.g. if you had 8-core CPUs but only allowed 6
jobs due to a lack of memory then you should only count the 6 cores
becase the others might as well not exist if nothing can use them.
However, it would then be fair to only run 6 copies of the benchmark.
Conversely, if you allow overcommitting, i.e. running more than 1 job
per core in some cases, you shouldn't advertise more LogicalCPUs than
cores unless you reduced HEPSPEC to compensate - but in general that
wouldn't be fair to you since most jobs would still in practice get the
full per-core power.
Stephen
--
Scanned by iCritical.
|