Hi All
As I said at the TB-SUPPORT meeting last week, ATLAS plan to run many
realistic user analysis jobs during the STEP09 tests in June. This is
an area where we have always been well below the potential capacity of
the grid, so we will be doing pre-testing from this week to ensure
that all of the infrastructure components are able to take the load
involved in running 17k simultaneous user analysis jobs.
To this end I have a request to the sites to:
1. Enable support for the /atlas/Role=pilot VOMS role in their batch systems.
2. Configure their batch systems to give the following shares to
different ATLAS activities:
- 50% to /atlas/Role=production (normal ATLAS production share)
- 25% to /atlas/Role=pilot (atlas analysis via panda)
- 25% to /atlas general (atlas analysis via WMS)
I know that creating new pool accounts and fiddling with LCMAPS is
tricky, but could sites please do this in the next couple of weeks? As
we do this at Glasgow I will try and follow up with more detailed
instructions for YAIM.
We are also trying to ramp-up the analysis tests so, to this end, we
plan to run a more intense version of hammer cloud tomorrow in the UK.
This will send up to 1200 jobs per site, which should keep the site
busy for most of the day running many analysis jobs.
We need ATLAS data to be present for sites to run this, so these would
be run tomorrow at:
UKI-LT2-RHUL_MCDISK
UKI-NORTHGRID-LANCS-HEP_MCDISK
UKI-NORTHGRID-LIV-HEP_MCDISK
UKI-NORTHGRID-SHEF-HEP_MCDISK
UKI-SCOTGRID-GLASGOW_MCDISK
UKI-SOUTHGRID-OX-HEP_MCDISK
UKI-SOUTHGRID-RALPP_MCDISK
UKI-LT2-QMUL_MCDISK
UKI-SOUTHGRID-CAM-HEP_MCDISK
UKI-SOUTHGRID-BHAM-HEP_MCDISK
If any site has a problem running tomorrow, please let me know ASAP.
We also plan to run a similar test at the end of the week using the
panda backend and user job pilots. These are "known" safe jobs, but
will obviously not use glexec uid switching, so if any site does not
want these jobs to run then, again, let me know.
Our tentative target is to achieve a rate of 10Hz in AOD analysis,
which is known to be challenging for the larger sites (see
http://tinyurl.com/d68hna and http://tinyurl.com/cfd5ww for some
tuning tips!).
I have one final question, which is about fairshare, does anyone
understand the difference between DEDICATEDPS and DEDICATEDPES in
maui? I want to make sure that fairshares are calculated based on
wallclock times, not on cpu consumption times. The maui manual
describes FSPOLICY (http://tinyurl.com/cjgpba), but doesn't really
explain which setting does what.
Cheers
Graeme
--
Dr Graeme Stewart http://www.physics.gla.ac.uk/~graeme/
Department of Physics and Astronomy, University of Glasgow, Scotland
DEATH TO MEETINGS!
|