A New STAR Event Reconstruction Chain

QuickTime™ and aTIFF (Uncompressed) decompressorare needed to see this picture.

A New STAR Event Reconstruction Chain

Claude A. Pruneau, M. Calderon, B. Hippolyte, J. Lauret, and A. Rose.

STAR Collaboration

Physics and Astronomy Department

Wayne State University

QuickTime™ and aTIFF (Uncompressed) decompressor

are needed to see this picture.

QuickTime™ and aTIFF (Uncompressed) decompressorare needed to see this picture.

QuickTime™ and aTIFF (Uncompressed) decompressor

are needed to see this picture.QuickTime™ and a

TIFF (Uncompressed) decompressorare needed to see this picture.

Claude A Pruneau, CHEP 2004 -2-QuickTime™ and aTIFF (Uncompressed) decompressorare needed to see this picture.

STAR Experiment• Multi-purpose detector for heavy ion and p-p physics• Multiple detector sub-systems


Challenges• Colliding systems: p+p to Au+Au collisions at 20 – 200 GeV/u.

• Large particle production– E.g. Few 10s of tracks (and pile-up) in pp to ~6000 tracks in central Au+Au

collisions, with up to 50 hits/track in the SSD, SVT, TPC detectors.

• Large kinematic range of interests/detection: 0.15 < pt < 20+ GeV/c; ||<4.

• Large range of physics analyses.

• Very large data volume; e.g. from Run 4: – 15106 Au+Au events @ 62 GeV.

– 94106 Au+Au events @ 200 GeV.

– 230106 p+p events @ 200 GeV.

– Raw data: 200 TBytes, DST: 40 TBytes.

• Evolving detector configuration: TPC, FTPC, SVT, SSD, FTPC, EMC, …


Data Analysis Challenge• Analysis proceeds (roughly) in two passes:

– Event Reconstruction - ideally one pass @ central facility – End-user physics analysis.

• Old reconstruction software – Separate track finding (TPT) and Kalman Fit (EGR)– Appropriate for TPC analysis.– Deployment of new detectors (e.g. SVT, SSD) required new

tracking modules in a patch work fashion.• Tracking time increases linearly with number of detectors rather than

hits.

– Mix of FORTRAN, C, C++ codes difficult to maintain.– Limited documentation – Developers no longer collaboration members– CPU Time/central event : 115 s.


Outline• Need for a new reconstruction chain • New tracker

– Description– Performance

• New Reconstruction chain– Summary of changes– Performance verification

• Summary


Goals of New Tracker• Develop/Deploy an integrated event reconstruction software/environment.

– Integrate track finding and Kalman fitting in one package. – Enable integration of existing and new detectors in a single analysis framework.

• Develop object models and algorithms that enable flexibility and growth. – Detector geometry representation– Hit and track representations.

• Match or Improve track reconstruction performance– Reconstruction efficiency– Resolution– Kinematic range/acceptance

• Eliminate legacy Fortran code.• Code Robustness

– Memory leak free.– Proper handling of unforeseen exceptions.– Reduce memory footprint.

• Reduce reconstruction time.• Provide abundant documentation.


Integrated Tracker Task Force (ITTF)

• Members:Manuel Calderon1, Jerome Lauret1, Lee Barnby2, Camelia Mironov2, Ben Norman2, Maria Mora Corral3, Mike Miller4, Zbigniew Chajecki5, Claude Pruneau6, Andrew Rose6

1 Brookhaven National Laboratory, 2 Kent State University, 3 Max Planck Institut fur Physik, 4Yale University,5 Warsaw University of Technology,6

Wayne State University.

• Formed : November 2000.

– Mandate: Design/Develop/Deploy a new, integrated tracker for STAR.

• Design/Development 2001-2002.

• Design Review : Sept 2002.


Tracker Design Considerations• Language/Portability/Minimal dependencies

– ANSI C++ & Standard Template Libraries (STL).• Modularity and Expandability

– Define/Use generic interfaces for key components– Plug-and-play Components and Algorithms.

• Algorithms and Object Models

– Non STAR detector specific.

– Simple Geometry Model (Detector, Shape, Placement).

– Elementary Constructs (Track, Hit, etc).

– Special Containers when appropriate – Detector Geometry, Hits

– Generic Algorithm/Interface

• Track finder and seed finder.

• MCS, E-loss, dE/dx Calculations

• Hit error parameterizations

– Templated Object Factory and Memory Management.

– Abstract Input/Control Parameter Representation.


Some Implementation Specifics• Detector Geometry Representation and Traversal

– Generic basic detector class. (Placement, shape representation, materials, hit error parameterization, energy loss calculation).

– Detector groups (TPC, SVT, …) and builders to instantiate & assemble geometry.

– Detector tree (e.g. sorted radially, azimuthally) for traversal

• Tracking algorithms (non detector specific): – Abstract interface notion of track finding.– Concrete classes for track seed finding & track finding/fitting.

• Hits: – Interface provides for access in local (detector) or global coordinates – Storage in tree/map for fast retrieval

• Hit loaders – Abstract interface define notion of hit loader (Generic).– Use one loader per hit bearing detector group (Star specific).

• Output to persistent STAR data model (StEvent).


Object Factory + Memory Management Model

• Double Template Class : class VectorizedFactory<T1,T2>– First template (T1): Class to actually instantiate.– Second template (T2): Base class served by the factory.– Use STL vector class for storage.– Memory allocation and garbage collection done in one place – Nominal set of object instantiated/destroyed once at startup/finish time. – Object set expanded in large blocks as needed.

• Pros:– Avoid repeated calls to "new" and "delete" for each event analysis.– Enables plug and play of new components.– Enables run time choice of classes to instantiate and use.– Simplified user code – no memory management.– No memory leaks. – Promotes code speed, simplicity, and robustness.

• Caveat: – Reused objects must be properly reset and initialized– Large footprint.


Tracking Algorithm • Kalman Filter/Finder• Local Helix Model• Local (detector) coordinates• Detector geometry integrated

– Multiple scattering, Energy Losses

Outside-in pass

Inside-out pass

Seed

Collision Vertex

rB Find seed

Outside-in Find/Fit Pass

Extend Track To Vertex

Extend outward?

Filter/Save track

Find Vertex

Reset Track Container

Load Hits

Inside-out Find/Fit Pass

Filter/Save track

Done


B

rX =

yz

=Cxo

C =1/ Rtanλ

⎡

⎣

⎢⎢⎢⎢⎢⎢

⎤

⎦

⎥⎥⎥⎥⎥⎥λ

Track model: Local helixx

Tracking Algorithm: Kalman Filter - Local detector frame.

(xo,yo)oCx=

(y, z) y

R

Prediction

Update


Kalman Tracking/Fitting

C(E ') =C(E) + ΔEdC(E)

dEE

=C(E) 1−p2 + m2

p2 ΔE⎡

⎣⎢⎢

⎤

⎦⎥⎥

x̂

k +1− = f x̂k( )

Pk +1− =FkPkFk

T + Qk

F

k=

∂rfk

∂rxk

rX =

yzC

tanλ

⎡

⎣

⎢⎢⎢⎢⎢⎢

⎤

⎦

⎥⎥⎥⎥⎥⎥

rZ

k=

yz

⎡

⎣⎢

⎤

⎦⎥

Hk=

1 0 0 0 00 1 0 0 0

⎡

⎣⎢

⎤

⎦⎥

K

k=Pk

−HkT HkPk

−HkT + Rk( )

−1

P

k= I −KkHk( )Pk

−

x̂

k=x̂k

− + Kk zk −Hkx̂k−( )

Compute Kalman Gain Kk

Prior Estimate xo-

Error covariance Po-

Update estimate with measurement zk.

Update Error covariance.Project ahead.

(Eloss Correction)

Find best matching hits


Δ

Δ

Δ

pt

Δ

pt

Primary +

ITTF REDTPT BLUE

Reconstruction Bias & Resolution


Δp

t/pt>

Δp

t/pt)

p(GeV/c)

Reconstruction Bias & ResolutionPrimary +

ITTF REDTPT BLUE

Δp

t/pt)

p(GeV/c)

Occupancy RED - LowBLUE - High


Efficiencies vs pt

Cuts : N MC Hit > 10; -1<<1; DCA<3 cm

pt (GeV/c)

ITTF

TPT

Low Multiplicity High Multiplicity

pt (GeV/c)


ITTF Review

• New Tracker Performance Review– STAR Internal Panel Review - Aug 2003.

– Found New Tracker to have equivalent or better performance

– Recommended adoption of the new software for integration in STAR data reconstruction production.

• Official Adoption of the new tracker by STAR– Aug 2003 Collaboration Meeting.


New Reconstruction Chain

• Goals: – Integrate the new (ITTF) tracker in the STAR reconstruction

chain.– Eliminate obsolete/legacy code; e.g. tables .– Use StEvent as object model both for processing and

persistency.

• Integration/Development Team– J. Balewski, M. Calderon, L. Didenko, Y. Fisyak, B.

Hippolyte, J. Lauret, M. Oldenburg, C. Pruneau, A. Rose.

• Duration: ~ 7 months.


Changes and Improvements…

• ITTF Track Reconstruction (ITTF Team)• Generic Vertex Maker (L. Barnby, J. Balewski, T. Ulrich)

– Façade/Interface deployed to enable multiple vertex finding algorithms. – New Maker based on Minuit

• TPC cluster finder (DAQ Team: J. Landgraf, T. Ljubicic )– Fast finder, re-written from scratch in C++.

• Kink finder (C. Mironov, S. Margetis)– K reconstruction– C++ re-write of FORTRAN code.

• TPC Hit Calibrations (J. Lauret)– Coordinate transformation, calibration adjustment) – Formerly entangled with old tracker TPT, now a new module

“StTpcHitMover” .

• Addition of chain options for increased flexibility (J. Lauret, Y. Fisyak, M. Calderon)

– Module ordering no longer static, but predicated on components included in the Chain.


Changes and Improvements…

• SVT Code– Modified to use StEvent (the persistent data model) or tables.

• FTPC Code – Modified to use StEvent.

• Trigger data detectors– Formerly in StEventMaker, now part of a compendium maker (“foure-tout”)

• Performance Evaluation Codes (J. Lauret, Y. Fisyak, M. Calderon)– Included propagation Geant particle ID in hit/track reconstruction to

dominant contributor evaluation (key to dominant, number of hits, avg quality).

– Generic track-track comparison maker was developed.

• Integration of SSD.

• R&D for a new pixel detector


Performance Verification• Goals

– Verify code integrity - produce sensible numbers– Verify physics performance

• Tester team– Representatives from each STAR Physics working group– Event structure - Aya Ishihara– Spin - Jan Balewski – HBT - Zbigniew Chajecki – Heavy Flavor - Alex Suaide – EbyE - Paul Sorensen – Spectra - Johan E. Gonzalez, Alexander Wetzler– High-pT - Marco van Leeuwen– Strangeness - Sevil Salur. Camelia Mironov, Ying Guo

• Duration: June 16, 04 to September 22, 04.• Hard work!

– Data samples reproduced 14 times. – Multiple (minor) bug fixes.

mailto:[email protected]



















Performance Verification

N

Balew

ski

d + Au @ 200 GeV.

NFitPoint15, DCA ≤ 3 cm, |≤1.00, pt 0.40 GeV/c, FitPtefrac 0.55, Zvert ≤ 100.

pt (GeV/c)

Yie

ld

Yie

ld

pt (GeV/c)

Mass (GeV/c2)

A W

etzle

r

Previous - Black

New - RED

S. S

alur

J. E. G

onzalez



Performance Review - HBT Analysis - Track Merging/Splittingd +Au data

Previous New

Zb

ign

iew

Ch

aje

cki

CF

Average Separation (cm)

Merging

Splitting

Zb

ign

iew

Ch

aje

cki


Previous New BETTER!!! Zbig

nie

w C

haje

cki


Summary

ITTF Tracker Completed!

Integration in Reconstruction Chain Completed!

Performance Comparison/Validation Completed!

Production of Run 4 data Started!


Integration of new detectorsWork in progress (K. Schweda, et al.)

Addition of SSD and R&D for a new pixel based vertex detector


Epilogue• ITTF originally conceived as a 2 years effort

– To be conducted by a handful of people.

• STAR is a successful on-going experiment.– Taking data, and publishing physics.

• Code development + deployment – Took quite a bit longer than anticipated.

– Required participation of very many people to establish code integrity, and for performance evaluation.


Worth the effort – Powerful, Flexible, Integrated Tracker

• Robust - Memory leak free, Good exception handling.• Faster. • Allows evolution

– Proven Performance• Performance comparable or better than that of old tracker.

• All STAR PWGs signed-up on its value based on performance

achievements in a wide spectra of analyses and level of details.

– Easy Maintainability.

– Documented.

– Ready for the Future decade!!!! • On time for STAR R&D developments

Documents

A New STAR Event Reconstruction Chain