Hi,
I took the opportunity this week to clarify a few things with ATLAS that
I thought it would be helpful to share.
Memory: ATLAS needs sites to provide 2GB of memory per job. This was
originally specified in 2004 and remains their position. They currently
require 1100MB per reconstruction job from the info system, but this
will be changed from time to time and they reckon 2GB will cover all
eventualities. This requirement still holds for 64-bit machines too.
Internal site bandwidth for analysis: the bottom line is a very rough
estimate of 2-4 MB/s (average), 10 MB/s (peak) per job. More info below.
Since no one can really know what typical ATLAS user analysis code will
be like yet, estimates are a bit speculative. Nonetheless, several
people have noticed that Root appears to have an intrinsic limit of
10MB/s when reading in files. Furthermore, it seems unlikely that this
will be a limiting factor given the size of analysis data per event and
the sort of programs users will run on this data. Estimates of the
actual bandwidth for analysis are more like 2 to 4 MB/s, so if you
specify worker nodes, networks and SRMs which can comfortably handle
this on average plus a peak of 10 MB/s/job you should not go far wrong.
ATLAS analysis will use rfio (and possibly other protocols in the
future, but not gridftp) which is good news for sites with WNs on a
private network - you can avoid rfio (but not ftp) traffic going through
the NAT - see Graeme's useful wiki entry on this:
http://www.gridpp.ac.uk/wiki/DPM_Dual_Homing
Despite being a member of ATLAS I'm not acting in any official capacity
hear, just reporting what I found out as a site admin.
Cheers,
Simon
|