Tracking and Vertexing

rucksackbulgeAI and Robotics

Dec 1, 2013 (3 years and 10 months ago)

60 views

ALICE HLT High Speed

Tracking and Vertexing

Real
-
Time 2010 Conference

Lisboa, May 25, 2010

Sergey Gorbunov

1,2

1

Frankfurt Institute for Advanced Studies, University of Frankfurt, Germany

2

Kirchhoff Institute for Physics, University of Heidelberg, Germany


2
/18

Real
-
Time 2010

Sergey Gorbunov, FIAS

HLT event reconstruction scheme (main trackers)

The TPC Sector Tracker is the most

complicated algorithm:




combinatorial search



fit mathematics





the reconstruction time is crucial


3
/18

Real
-
Time 2010

Sergey Gorbunov, FIAS

HLT TPC Sector Tracker: The Cellular Automaton method

row k+1

row k

row k
-
1

TPC sector

1.
Neighbours finder:



For each TPC cluster it finds
two (up&down) neighbours which
compose the best line

2. Evolution



non
-
reciprocal links removed



one
-
to
-
one linked clusters are
compose track segments

3. Other steps



fit, search for missed
hits, and the final track
selection

4
/18

Real
-
Time 2010

Sergey Gorbunov, FIAS

HLT TPC Tracker performance (Sector Tracker + Global Merger) on MC

pp,

HLT

pp,

Offline

central PbPb,

HLT

central PbPb,

Offline

99.86%

9.06%

0.19%

98.15%

13.22%


1.66%

95.84%

12.13%


1.40%

99.94%

9.30%

0.21%

5
/18

Real
-
Time 2010

Sergey Gorbunov, FIAS

HLT TPC Tracker performance (Sector Tracker + Global Merger) on MC

HLT Tracker

:


Time =
17.6

s

Eff = 98.15%

Ghost = 1.66%

Clone = 13.22%

HLT Tracker

:


Time =
19.6

ms

Eff = 99.86%

Ghost = 0.19%

Clone = 9.06%

Offline Tracker

:


Time =
160.1

s

Eff = 95.84%

Ghost = 1.40%

Clone = 12.13%

Offline Tracker

:


Time =
66.0

ms

Eff = 99.94%

Ghost = 0.21%

Clone = 9.30%

MC, 14 TeV pp events:


MC, 5 TeV Central PbPb events:

Performance on Monte Carlo

~ linear time dependence:

6
/18

Real
-
Time 2010

Sergey Gorbunov, FIAS

Real PP Event in the HLT (2009 data)

primary vertex

vertex
-
fitted tracks

tracks

7
/18

Real
-
Time 2010

Sergey Gorbunov, FIAS

HLT V0’s: PP run 00010480

Monitoring of V0 physics on
-
line



HLT V0 finder



Gamma, Ks, Lambda analysis

8
/18

Real
-
Time 2010

Sergey Gorbunov, FIAS

Use of parallel hardware: GPU devices

NVIDIA GeForce GTX 280:



30x8 general propose processors;

pure calculations can be ~100 times

faster than CPU



very parallel: || execution of branches,
|| memory access



CUDA language
-

a little extension of
C++



fast access to the small portion of data
(16k) at the time; no memory cache



single precision floating point



ONLY parallel calculations

9
/18

Real
-
Time 2010

Sergey Gorbunov, FIAS

Running the sector tracker on the GPU cluster at Frankfurt University

CPU

GPU



speed
-
up: 10.5x



same code



same result

CPU

GPU

10
/18

Real
-
Time 2010

Sergey Gorbunov, FIAS

HLT Tracker
-

Summary

HLT Tracker
-

Summary:




The ALICE HLT tracker shows good performance and speed.




It is able to use the GPU hardware, showing ~10 times speed
-
up with
comparison to CPU.




The HLT was running well in 2009 performing the full on
-
line event
reconstruction, which includes monitoring of the events, the vertex position,
and v0 physics.

In work:




Installing the GPU hardware.



Further speed
-
up of the tracker.



Speed
-
up of the rest of the HLT reconstruction software for
heavy ions (clusterfinders, vertex finders, v0, ITS tracker, …).

11
/18

Real
-
Time 2010

Sergey Gorbunov, FIAS

HLT vertexer: the Silicon Pixel Detector

Silicon Pixel Detector (SPD)



The innermost ALICE detector



Two layers of silicon at ~4cm and ~8cm



Pixel measurements (XYZ)

The SPD detector can provide stand
-
alone

event vertex, which is useful for on
-
line

monitoring of the ALICE interaction point.

High
-
Speed Vertexing in HLT

12
/18

Real
-
Time 2010

Sergey Gorbunov, FIAS

HLT SPD vertexer: tracks



all combinations of inner + outer pixels



straight trajectories: the magnetic field is not taken into account

outer layer

inner layer

pixels

SPD tracks:

13
/18

Real
-
Time 2010

Sergey Gorbunov, FIAS

HLT SPD vertexer: track selection

1.
Calculate DCA point for each track.

2.
Remove tracks with DCA > R cut.

3.
Store Z of the DCA point in an array.

4.
Find the highest peak in the Z
-
array,


select tracks which produce the peak.

vertex guess

R cut

DCA point

Selection of tracks: cut in XY, search in Z

14
/18

Real
-
Time 2010

Sergey Gorbunov, FIAS

HLT SPD vertexer: vertex fit

The 3D vertex is fitted as a closest point to all the selected tracks.

15
/18

Real
-
Time 2010

Sergey Gorbunov, FIAS

HLT SPD vertexer: complete algorithm

1.
Guess the vertex

2.
Select tracks for the vertex fit

3.
Fit the vertex

4.
Iterate from step 2. several times

Complete algorithm:

16
/18

Real
-
Time 2010

Sergey Gorbunov, FIAS

HLT SPD Vertexer performance on MC

HLT SPD Vertexer performance



1000 14TeV pp MC events



In order to check possible bias to the origin, the MC vertex is set to (.2,.2,0)

Resolutions:



X,Y = 269 um



Z = 164 um



No offsets

Speed:

3500 events / s

17
/18

Real
-
Time 2010

Sergey Gorbunov, FIAS

HLT SPD Vertexer: real data 2009

On
-
line SPD vertexer on the HLT event display, run 101498

18
/18

Real
-
Time 2010

Sergey Gorbunov, FIAS

Summary




On
-
line SPD vertexer has been developed for the ALICE HLT.



It is fast and shows good resolutions on Monte
-
Carlo.



The vertexer is used for on
-
line monitoring of the ALICE interaction point
since the first collisions in 2009.

HLT Vertexer


Summary: