Testbed Support for GridPP member institutes
> [mailto:[log in to unmask]] On Behalf Of Winnie Lacesso said:
> BTW what are the right numbers for
> CE_PHYSCPU=
> CE_LOGCPU=
> when one gets only a percentage of a shared HPC?
I'd say that it depends on whether you have a hard limit or not. If the
grid has a definite maximum usage then publish that, but otherwise I'd
suggest publishing the full system, otherwise you may have reported
usage greater than your published capacity which would look odd. You are
also supposed to publish the VO shares, at least for the LHC VOs, so
whatever you do that should correspond - so e.g. your installed capacity
for CMS will be calculated as cms_share * LogicalCPUs * HEPSPEC.
> (Pointer to lucid online policy documentation welcome)
Well, just the standard "Flavia" document - you can send her any
requests for clarification, I think she's collecting a list!
> As Jon/Yves configured it (which I inherited), they quote
> all the HPC WN resources.
> So Bristol shows as red on some monitoring pages, since there
> are many
> free jobslots while also many PP-gridpp jobs queued.
The free job slots count is separate from the logical/physical CPU
numbers. The former come from the dynamic info providers, which may or
may not be doing what you want - they should count the number of jobs
which could be started, at least potentially. Anyway in general it isn't
an error to have jobs queued and slots free because it isn't possible to
reflect all the detail of the scheduler, e.g. one user may have queued
10,000 jobs but probably won't be allowed to fill the system ...
Stephen
--
Scanned by iCritical.
|