Core Minutes 2/22/2005
Present: Pol d'Avezac, Joanne Bogart, Anders Borgland, Toby Burnett,
James Chiang, Johann Cohen-Tanugi, Seth Digel, Richard Dubois,
Zachary Fewtrell, Riccardo Giannitrapani, Berrie Giebels,
Navid Golpayegani, Tony Johnson, Heather Kelly, Matt Langston, Chuck Patterson,
Leon Rochester, Tom Stephens, Tracy Usher
- Data Handling: (Tony)
- System Tests No news since the release of a new
version last week.
- Pipeline Dan was sick last week; no news.
- Data Server Planning a meeting with Jean-Paul Lefevre.
- GlastRelease: (Richard) Focus is on
making a GlastRelease tag which will form a suitable base for EM
tag which I&T can use with the new hardware being assembled.
(Toby) has been making new tags to incorporate the latest and
greatest Tracker code, but backed out some CalRecon changes since the plan is
to separate this upgrade from the Tracker upgrade.
He will ask Julie to run System Tests
on the most recent, v6r2 (but see note below about clone of
CalValsTool). (Richard) Still some concern on the following fronts:
memory leaks, increased processing time, and significant differences
in output (e.g., number of tracks). In the new calorimeter code, CalDigi takes
about 20 times longer than it used to, about 1/3 second per event for 100 MeV
events. First stage of TkrRecon may also be somewhat slower.
- Tracker:
Tracy has been tracking down memory leaks. After switching to
a newer, better version of ValGrind (thanks to Anders for the suggestion),
he found three new leaks in TkrRecon. He also fixed a problem causing
very long processing time in those (infrequent) events where a particle was
near a boundary and had a trajectory making a shallow angle with the boundary.
He noted there have been changes to the combinatoric pattern recognition,
causing it
to find more tracks, so it is not surprising that processing time in
this stage would increase, or that results would be somewhat different.
Meanwhile Leon has been busy on several fronts:
- A week ago he introduced a bug into TkrUtil while making changes to
accommodate, Tower A, which had no CAL. It broke code that calculated
CAL volume identifiers when there was a CAL. This is now fixed
in the latest tag of TkrUtil.
- There is a clone of the (AnalysisNtuple)
CalValsTool class in CalRecon, presumably to avoid dependence on
AnalysisNtuple. However changes needed for
new TkrRecon were not completed on this clone.
Leon has finished the job and committed
the changes to CVS, but will leave the task of tagging to someone more
versed in the current state of CalRecon branches.
- In order to see if the modified pattern recognition logic was affecting
processing time, he tried backing it out (on his Windows laptop).
Didn't appear to make much
difference.
- Using a GUI event display (the old one or FRED) makes a huge difference
in per-event memory leaking. Without a GUI, after all Tracy's work, we're
down to about 800 bytes/event. With a GUI, it's more like 100 Kbytes!
- Calorimeter:
Berrie reported that a first check on CalRecon with 10 GeV photons
generated by the system tests on v6r1p2
shows results consistent with those of v4r5p1. Compare for example this
v4r5p1 plot
with the v6r1p2
version.
We applauded and Berrie bowed..
- Readiness for I&T:
(Anders) I&T would like to get the new TkrRecon
into EM as soon as possible, but it is clear
there are issues. They will discuss the options today.
Some data may be taken today.
(Richard) What version of EM is currently being used? Answer:
v3r0407p3. They plan to wait until there
is a more useful tag available, rather than jumping on p10. (Heather)
p11 will be made available and
will include the new GEM; then there will be a new tag of EM for
the new TkrRecon. (Leon) He has new calibration code on his laptop.
There is a remaining defect in that missing strips
can result in broken tracks. The point is that there
will still be some unsatisfactory aspects to the
new code...not mysterious, but it will need work.
Kalman energy is also a work in progress.
(Richard) We all knew what we were getting into
when we signed up to provide new code to I&T -
chins up..this is to be expected.
- RM: (Navid) Concerning the Windows version, he
saw in CMT source code that it uses _chdir, which can only handle paths of the
form: X:\blah, meaning samba mounts would have
to be mapped to a windows drive...But LSF on
windows does not accept network mapped drives.
Work continues to sort out the cause of the slowness when accessing
the RM over the web. You can see from
this plot that
web queries via the production SLAC server are always somewhat slower than
web queries via glast02 (which are uniformly very fast); often they
are much slower. As Navid said in an accompanying mail message,
"web query on www.slac.stanford.edu from noric05" is the most interesting
one. I assume it's a fairly busy server, so it's normal for the wall
clock time of the cgi script to vary more than the other 2. What's
unusual, though, is that it's varying by such a big amount. Some queries
do run in reasonable time (<1sec) while others are taking a long time.
There are a few outliers in there that I cut off from the graph. Some
queries on www.slac.stanford.edu took up to 100 seconds to reply.
-
Personnel: (Richard) told Human Resources to
stop accepting applicants for the Alex replacement. We plan to make an offer
to David Brock; after this meeting we have one last interview with
Samuel Johnson.
Action Items
- Toby will branch or retag CalRecon (Berrie says ok) to put in
the repaired CalValsTools clone.
- Tracy & Leon will continue looking at TKR to see if they can
find plausible cause for the increase in processing time.
- Zach will track down cause of increases in processing for CalDigi.
J. Bogart & H. Kelly, Last Modified:
01-Jun-2010 15:45:00 -0700