Fast Pattern Recognition for CMS Track Finding

1
Fast Pattern Recognition for CMS Track Finding

• Malina Kirn
• 9/28/2006
• Project Proposal, AMSC 663
• Advisor Dr. Nicholas Hadley

2
Outline

• Goal Improve the running time of an existing
  track finding software package designed for the
  CMS experiment while preserving physics
  performance.
• Science, CMS, Tracker
• Track finding algorithm
• Project

3
Science Goals

• Higgs Boson
• New Physics
• Dark Matter

4
(No Transcript)
5
(No Transcript)
6
(No Transcript)
7
(No Transcript)
8
(No Transcript)
9
Si strip detectors
Tracking Detectors
Pixels
10
Problem Size
Pixels x 66 million
100 µm
150 µm
Stripsx 9.6 million
10-25 cm

• Final data rate of 150 Hz
• 1,000-10,000 hits/event

80-180 µm
11
double sided strip sensor (stereo)
single sided strip sensor
TOB
r
TIB
TID
TEC
z
12
The RoadSearch Algorithm

• Goal reconstruct tracks from hits
• Use roads to reduce the problem space

13
The RoadSearch Algorithm

• Create seeds trajectories
• Seed (ideally) the end points of a track
• Trajectory first guess track
• Create clouds
• Cloud clustering of hits around trajectory
• Clean/merge clouds
• Find track candidates
• Create tracks

14
Roads
15
outer Rings for RoadSeeds
inner Rings for RoadSeeds
r
z
16
1) Create seeds trajectories

• Match pairs of hits in the inner and outer rings
  of a single road.
• Create seed if pair satisfies ?f cut.
• Create trajectory using the three points of beam
  spot and inner/outer hits.
• One real track can have multiple seeds.
• One event typically contains multiple seeds and
  trajectories.
• Goal identify tracks

17
1) Create seeds trajectories
r-z plane
r-f plane
18
2) Create clouds

• Loop over rings in road from inside?outside.
• Include hits in a ?f window, which gets larger
  with r.
• Cloud contains one trajectory and multiple hits.
• Goal identify hits associated with track

19
3) Clean/merge clouds

• Can have multiple clouds/track because of
  possible multiple seeds/track
• A single track described by multiple clouds is
  bad
• Merge clouds which share 70 of their hits.
• Create new trajectory based on merged hits.
• Cloud contains one trajectory and multiple hits.
• Goal one cloud per track

20
4) Find track candidates

• Create one track candidate/cloud.
• Start with trajectory.
• Add hits from inside?out if ?2 from new hit is
  less than cut.
• Update the track candidate.
• Pass a cloud to (5) with the track candidate and
  only good hits.
• Goal keep only best hits, create initial track

21
5) Create tracks

• Given track candidate good hits, create one
  track.
• Iteratively apply Kalman filter on all hits.
• Goal final track creation

22
Tracks!
23
Project

• Final data rate of 150 Hz
• Event reconstruction at CERN farm (order 1000
  cores) must be 150 Hz
• CMS outputs 5 PB/year
• Roadsearch is currently too slow
• Track finding consumes 50 of time
• Plan algorithmic and technical improvements

24
Project Platform

• CMS uses several grids for data distribution
• Must be able to test and run at any grid site
• Software installed at all grid sites
• Constrained to run within software
• Cannot make hardware-specific improvements
• I will develop at Fermilab

25
Project Test Problem

• Monte Carlo generated events
• Use MC single and multiple track events to
  measure success rate and running time
• Speed improvements desirable for all event types
• Must know speed success rate before after
  changes
• Comparison to other tracking algorithm desirable

26
Summary

• Roadsearch is a new track finding algorithm with
  unique features
• Goal is to improve Roadsearch running time
• Grid puts constraints on possible solutions
• Testing done on well-understood datasets