Testbed Support for GridPP member institutes
> [mailto:[log in to unmask]] On Behalf Of John Gordon said:
> Alessandra, the approximation of 4 is just to allow a rough comparison
> between sites which are doing their accounting based on SI2K and those
> who have measured HEPSPEC06.
There are also a couple of practical points. One is that GLUE 1.3 only
has SI00 (and SF00) as benchmarks, and people may use them e.g. in their
JDL to estimate what CPU time limit they need, so we wanted to keep
publishing a number with roughly the same meaning - if values suddenly
stepped from 2000 to 8 it could cause problems (and the attribute is an
integer so we would lose accuracy). OTOH if they step by 10% that
shouldn't be a problem as any users should be allowing a headroom
significantly bigger than that as benchmarks are only a rough guide
anyway.
Conversely we need a single common scale factor since we want to be
able to interpret the SI00 attribute as HEPSPEC across all the sites
which are using it. We could have picked a conversion factor of, say,
4.2835, but individual sites would still have varied widely around that
and it would have been hard to remember and even harder to calculate in
your head!
The other thing is that, as far as I know, this is only an LCG thing,
so sites which haven't pledged resources to LCG (even if they support
LCG VOs) aren't bound to do it and the EGEE infrastructure in general
needs to continue to support the old method. John may know more but I
haven't seen any discussion of this in EGEE up to now, and I would think
there may be some resistance to adopting a benchmark with "HEP" in the
name ...
Stephen
--
Scanned by iCritical.
|