1. Status of the Offline Pattern Recognition Code Goa, March 12th 2013
Gianluigi Boca
GSI & Pavia University

2. Outline
Progress in the Offline Pattern Recognition since the last December
meeting : the central tracker code;
status of the V0 Pattern recognition;
a reminder to the Collaboration : need for a ‘time’ task for the
offline (and online also).
Gianluigi Boca, 12 March 2013

3. Present scheme of the offline PR for the central trackers
In the present PR there are essentially 3 steps :
1) do clusterization with STT axial + SciTil hits only, then do the fit
in the XY plane; find track candidate;
2) attach to track candidate Mvd hits that are near, redo the fit in
the XY plane;
3) attach skew Stt hits, do the fit in FZ and find remaining Helix
parameters;
Gianluigi Boca, 12 March 2013

4. Improvements

5. Improvements since last December meeting
improvement in the way clusterization with STT axial + SciTil
hits only is done;
2) improvement in the way the Mvd hits are attached to a track
candidate ;
Gianluigi Boca, 12 March 2013

6. 1) improvement in clasterization with STT axial + SciTil hits
Presently each Stt axial hit is considered as possible cluser seed (unles
already included in a previous track). For instance in this event the
cluster finding - tracklet finding is performed 110 times :
110 times
MC truth
reco track
Mvd Pixel
Mvd Strip
Stt Parallel
SciTil
event : 1 track, P = 0.3 GeV/c + BKG

7. 13 times Change the way the cluster search in XY plane is performed
Improvement : only Stt axial hits on the geometrical boundary of the
axial Stt region are considered as possible cluster seed.
13 times
MC truth
reco track
Mvd Pixel
Mvd Strip
Stt Parallel
SciTil
Gianluigi Boca, 12 March 2013

8. Before : require the Mvd hits near (within 5 mm) the circular
2) improvement in the way the Mvd hits are attached to
a track candidate :
Before : require the Mvd hits near (within 5 mm) the circular
trajectory in XY plane AND all attached hits must
belong to a Mvd Tracklet found previously with the
Mvd standalone Riemann Pattern Recognition;
Now : require the Mvd hits near (within 5 mm) the circular
trajectory in XY plane;
SAVE the Cpu time necessary for the Mvd
standalone Riemann PR.
Gianluigi Boca, 12 March 2013

9. Cpu time consumption

10. Paris meeting Cpu time performance of the code now
Cpu times measured on Intel Xeon 2.13 GHz 64 bit Lenny machine
MC generation : Box Generator;
multiplicities from 1 up to 8; momenta from 0.3 to 10 GeV/c.
Also I discover an error in the Cpu time calculated in my previous presentations, for the events with pileup mixing. The Cpu time
consumed by the task that produces the pileup of the hits was added in by mistake.
Paris meeting
So, we started off in Paris with 9.5 sec/evt for evts with pileup mixing …….

11. …… and now we are at 18 msec /evt , a factor 500 better
No Bkg
Bkg added
But this is not the end of the story because more improvements can be
achieved …….

12. ….. possible further improvements in speed :
1) confirm track with a SciTil hit at the earliest possible stage, not
done yet  save time by rejecting ghost tracks early;
2) eliminate some redundancies still present in the algorithm;
3) parallelize the algorithm;
[ 4) …… measure the Cpu time on a more modern computer ,
this alone probably gives a gain of a factor 1.5 – 2 ]
Gianluigi Boca, 12 March 2013

13. Even without further likely improvements we are in
the right ballpark with the Cpu time :
1) suppose a (conservative) factor 1000 coming from the multicores
available for Panda ;
2) suppose true the factor 1000 coming from an online trigger;
if all that’s true 18 msec/evt translates in 18 nsec/evt effective Cpu
time.
Gianluigi Boca, 12 March 2013

14. …but in any case, the work for improving this piece
of code continues, here is the more immediate
road map :
1) Check track efficiencies ( I didn’t have time for this meeting);
2) confirm track with a SciTil hit at the earliest possible stage, not
done yet  save time by rejecting ghost tracks early;
3) eliminate some redundancies still present in the algorithm;
4) refine the Cleanup code;
5) parallelize the algorithm;
6) finalize the V0 package.
Gianluigi Boca, 12 March 2013

15. Status of the V0 Pattern Recognition
(by Lia Lavezzi)

16. New procedure
Up to now the secondary track finder was a sort of adaptation of the primary track finder to tracks not originating from (0, 0, 0)
The impossibility to apply the constraint of being primary creates problems:
the conformal map transforms circles in straight lines only if the circles pass in (0, 0);
the fit works fine only with a big lever arm, so without the (0, 0) point it might give underestimated radius circles (with GLPK)
New implementation:
make a translation of the center on a precise hit, such as an MVD hit, a SciTil hit or the center of the STT in case of small isochrone;
apply the Legendre transformation and search for the peak in the accumulation histogram
Lia Lavezzi, 12 March 2013

17. Conformal Transformation
To obtain straight lines, a conformal transformation is performed
… BUT…
the track must stem from the origin of the coordinates
 a suitable point must be chosen as new origin
a translation is performed
points (x, y)
circles (x, y, rd)

18. Legendre Transformation
Alexopoulos T. et al, NIM A 592 (2008) 456– 462
For each hit we get two sinusoidal curve
The accumulation plot is filled and the peak. corresponding to a track, is found

19. Legendre Transformation
Preliminary Results with 1000 muon track events from IP
Reconstructable MC Track (> 3 STT // hits)
Reconstructed
Good (>80% of right hits assigned)
Wrong (< 80% of right hits assigned)
Not Reconstructed (Missed)
# Tracks
p (GeV/c)
Good
Wrong
Missed 1
0.3
0.91
0.07
0.02 4
0.94
0.03
0.04
0.88
0.06

20. A reminder to the collaboration

21. Need for a ‘time’ task or class for the Offline and
Online Pattern Recognition
All the Pattern Recognition code need to know the time when an
event occurred because from that it can be determined the time
window to consider for the hits to analyze.
Most of the times the SciTil detector can give that time with an
‘online’ imprecision of ~ 1 nsec [determined by the trajectory
variations of the charged particle producing the hit in SciTil].
This imprecision can be corrected later offline.
The time of the event can also be determined by the ‘triplet’ method
with the Stt especially for those events without tracks hitting the SciTil
system.
Gianluigi Boca, 12 March 2013

22. Need for a ‘time’ task or class for the Offline and
Online Pattern Recognition
I think we need a task or class (…or whatever you want to call it)
that delivers :
1- an event time based on SciTil signals and their corresponding
Stt triplet signals (weighted average of the two ? );
2- an event time based on the Stt triplet signals when there were
no tracks hitting the SciTil system.
I believe this is a delicate job that requires a dedicated person. What
does the Collaboration think ?
Gianluigi Boca, 12 March 2013