Java Reconstruction and Analysis for a Linear Collider Detector

Tony Johnson

Fermilab/Cornell Tutorial

January 2000

Contents

The Java hep.lcd framework for LC physics studies

Overview

Why Java?

Fast MC

Tracking Reconstruction

Cluster Finding

Java Analysis Studio

Distributed Physics Analysis

Java Performance

Conclusions - How to try it out!

Slide 3

What is hep.lcd?

A Java Framework for

Running Reconstruction and/or FastMC

Analysis of LCD data from Stdhep, Gismo, FastMC

A Tool for

Rapid Development of Reconstruction Algorithms

A Suite of Physics Analysis Tools

Histograming, Event Display, Jet Finding, etc.

Emphasis on Flexibility and Extensibility

Typically provide multiple reconstruction algorithms

Support for Large + Small + ... Detectors

Can be used standalone or inside Java Analysis Studio

Why Java?

Modern Object-Oriented language

Easy to learn and use

No need for backwards compatibility with C etc.

No pointers, memory leaks

Syntax very similar to C++, but without many of the more obscure, less useful features

Very Suitable for Rapid Prototyping

Powerful built-in utility libraries

Fast compiler, no link step, dynamic loading

No problems with porting code to new platforms

Can use Java Analysis Studio for interactive analysis

Bottom Line

If you know C++ Java will be familiar and a refreshing change

If you don’t yet know C++ Java is an excellent stepping stone

Either way you’ll have fun doing physics, not debugging crashes!

Java package hep.lcd

Reconstruction Processors

Track finder + track fitter

Several clustering algorithms

Parameterized MC Processors

Can read generator output (StdHEP) or Gismo output

Track and Cluster smearing

Analysis Utilities

Event Shape + Thrust utilities

Jet finders [Jade, Durham]

Histograming

Event Display

Simple 2D Event display currently

Track Finding/Fitting

Track Reconstruction

Track Finding uses M.Ronan’s (TPC) pattern finding

Tuned for Large + Small detector

Track Fitters:

SLD Weight Matrix Fitter

Can do Single Detector or Combined fit (e.g. VTX+TPC)

Kalman Filter (using FNAL MCFast fitter) coming soon(ish)

Fortran not Java, will need native library for each platform

Hit Smearing/Efficiency (since Gismo gives “perfect” hits)

Random Background overlay

What’s still needed:

More Track Finding Algorithms (Cheater, Projective Geometry)

End Cap tracking, Hit Merging, Realistic background overlays

Cluster Finding

Three Clustering Algorithms Currently Implemented

Cluster Cheater (uses MC truth to “cheat”)

Simple Cluster Builder (Touching Cells)

Radial Cluster Builder

All algorithms tend to produce many very low energy clusters - important to set sensible thresholds

Still Needed - Cluster Refinement Stage

Combine HAD + EM clusters

Endcap + Barrel overlap region

In Progress - Track Cluster Association

Initial Implementation Done by Mike Ronan

Output Format defined by Gary Bower

Need to Extend Definition of Clusters

Directionality, Entry point to calorimeter

Fast MC

Simple parameterized MC

Allows analysis directly from generator output without using full Gismo simulation

Produces same event format as Gismo

same analysis can be run with FastMC or Gismo. Can read StdHEP (generator files) or Gismo output

Produces tracks from MC particles based on parameters provided by Bruce Schumm

Produces Tracks + Error Matrices

Output compatible with Output from Full Recon

No Calorimeter Simulation at Present

Physics Utilities

Physics Utilities

4-vector, 3-vector classes

Event shape/Thrust finder

Jet Finder

Jade and Durham algorithms implemented

Extensible to allow implementation of other algorithms

Histograming (from Java Analysis Studio)

Event Display

Suitable for debugging reconstruction and analysis

Plan to use Wired for full 3D support in future

Particle Hierarchy Display

Contrib Area

Analysis Utilities and sample analyses provided by users

Event Display

Documentation and Tutorial

http://www-sldnt.slac.stanford.edu/jas/documentation/lcd/

Access to Code - CVS

Code recently moved to CVS for universal access

Browse CVS repository on Web

Connect with you favorite CVS client

protocol: pserver

server: sldl1.slac.stanford.edu

cvsroot: /nfs/slac/g/jas/lcdroot

userid: anoncvs

password: jascvs

module: lcd

Platform independent make (jmk) now used.

Most development currently done on NT

Now Unix development should be easy too

Java Analysis Studio

Set of experiment independent analysis tools for event oriented (High Energy Physics) data

Data Access classes provide access to many common HEP data formats

Histogram/Scatterplot Accumulation + Manipulation Classes

Plot Display classes

Lightweight framework for users to create physics analysis applications in Java.

Tools work alone, in combination, or within

Java Analysis Studio GUI which gives:

Integrated editor and compiler

Efficient access to local and remote data

Extensibility via Plug-ins, Fitters, Functions etc

GUI makes getting started easy
“Wizards” guide beginners

Built in Editor and Compiler
for writing analysis code

Histogram and Scatterplot display
Interactive Fitting and Rebinning

GUI can be extended to add
experiment specific features

Distributed Data Analysis with JAS

With many different simulated detectors and many physics processes, total MC data sample is large

JAS has built in support for efficient distributed physics analysis

Data Repository at Penn

LCD has set up central data repository at University of Pennsylvania (Don Benton), accessible from anywhere

Is Java fast Enough for HEP offline?

Current (266Mhz PII, JDK 1.1.7)

Clustering .6 secs/event

13.5 Million Calorimeter Cells

Fast MC 6 ms/event

Track Finding + Fitting ~5secs/event

Very competitive with C++/Root implementation (where they exist)

Will get even better!!!

JDK 1.2, HotSpot - Run-time optimization

In real life may be faster than C++

Better, cheaper performance analysis tools

Java encourages lightweight, module interfaces which promote efficient coding styles

People time is what really matters

Reconstruction Speed

Example of Using Track Recon.

Full Scale Analysis Possible (M. Ronan)

To Do List

Finish integration of MCFast Kalman filter

Support for merging signal/backgrounds

Additional Track Finders (projective, “cheater”)

Improved Cluster Description

Track/Cluster Association

Cluster Refinement

Vertex Finding code (based on ZVTop?)

WIRED Event Display

Support for SIO format reading/writing

Switch to XML based geometry description

Small Angle Tracking

Reference Analyses

Try it out!

Come to the tutorial this afternoon

Works on Windows (95/98/NT] or Unix (Linux, Solaris,…) or Other

Online tutorial available

Suitable for complete beginners:

no knowledge of Java or JAS assumed

starts with instructions on downloading and installing

Shows simple sample analysis jobs

http://www-sldnt.slac.stanford.edu/jas/documentation/lcd/

JAS Home Page

http://www-sldnt.slac.stanford.edu/jas

New Features in JAS 2.0

Many bug fixes - Printing Works - Save Histograms as GIF’s

Easier to install LCD extensions

no need to edit jas.ini

Support for logging output to a file

Supports 2D Binned Histograms (in addition to 1D Histograms and Scatter Plots)

Standalone jobs can save histograms (for later viewing in JAS)

Local jobs can append multiple datasets

Built-in LCD documentation/help

Gotcha’s for LCD users

Must use JAS 20 version of lcd.jar (in lcd download area)

Cannot connect to old servers (SLDNT0 OK)

JAS 2.0 – Now available from JAS website


	The Java hep.lcd framework for LC physics studies
		Overview
		Why Java?
		Fast MC
		Tracking Reconstruction
		Cluster Finding
	Java Analysis Studio
	Distributed Physics Analysis
	Java Performance
	Conclusions - How to try it out!


	A Java Framework for
		Running Reconstruction and/or FastMC
		Analysis of LCD data from Stdhep, Gismo, FastMC
	A Tool for
		Rapid Development of Reconstruction Algorithms
	A Suite of Physics Analysis Tools
		Histograming, Event Display, Jet Finding, etc.
	Emphasis on Flexibility and Extensibility
		Typically provide multiple reconstruction algorithms
		Support for Large + Small + ... Detectors
	Can be used standalone or inside Java Analysis Studio


Modern Object-Oriented language
Easy to learn and use
	No need for backwards compatibility with C etc.
	No pointers, memory leaks
	Syntax very similar to C++, but without many of the more obscure, less useful features
Very Suitable for Rapid Prototyping
	Powerful built-in utility libraries
	Fast compiler, no link step, dynamic loading
No problems with porting code to new platforms
Can use Java Analysis Studio for interactive analysis
Bottom Line
	If you know C++ Java will be familiar and a refreshing change
	If you don’t yet know C++ Java is an excellent stepping stone
		Either way you’ll have fun doing physics, not debugging crashes!


	Reconstruction Processors
		Track finder + track fitter
		Several clustering algorithms
	Parameterized MC Processors
		Can read generator output (StdHEP) or Gismo output
		Track and Cluster smearing
	Analysis Utilities
		Event Shape + Thrust utilities
		Jet finders [Jade, Durham]
		Histograming
	Event Display
		Simple 2D Event display currently


Track Reconstruction
	Track Finding uses M.Ronan’s (TPC) pattern finding
		Tuned for Large + Small detector
	Track Fitters:
		SLD Weight Matrix Fitter
		Can do Single Detector or Combined fit (e.g. VTX+TPC)
		Kalman Filter (using FNAL MCFast fitter) coming soon(ish)
			Fortran not Java, will need native library for each platform
Hit Smearing/Efficiency (since Gismo gives “perfect” hits)
Random Background overlay
What’s still needed:
	More Track Finding Algorithms (Cheater, Projective Geometry)
	End Cap tracking, Hit Merging, Realistic background overlays


Three Clustering Algorithms Currently Implemented
	Cluster Cheater (uses MC truth to “cheat”)
	Simple Cluster Builder (Touching Cells)
	Radial Cluster Builder
		All algorithms tend to produce many very low energy clusters - important to set sensible thresholds
Still Needed - Cluster Refinement Stage
	Combine HAD + EM clusters
	Endcap + Barrel overlap region
In Progress - Track Cluster Association
	Initial Implementation Done by Mike Ronan
	Output Format defined by Gary Bower
	Need to Extend Definition of Clusters
		Directionality, Entry point to calorimeter


Simple parameterized MC
	Allows analysis directly from generator output without using full Gismo simulation
	Produces same event format as Gismo
		same analysis can be run with FastMC or Gismo. Can read StdHEP (generator files) or Gismo output
Produces tracks from MC particles based on parameters provided by Bruce Schumm
	Produces Tracks + Error Matrices
	Output compatible with Output from Full Recon
No Calorimeter Simulation at Present


Physics Utilities
	4-vector, 3-vector classes
	Event shape/Thrust finder
	Jet Finder
		Jade and Durham algorithms implemented
		Extensible to allow implementation of other algorithms
Histograming (from Java Analysis Studio)
Event Display
	Suitable for debugging reconstruction and analysis
	Plan to use Wired for full 3D support in future
Particle Hierarchy Display
Contrib Area
	Analysis Utilities and sample analyses provided by users


Code recently moved to CVS for universal access
	Browse CVS repository on Web
	Connect with you favorite CVS client
		protocol: pserver
		server: sldl1.slac.stanford.edu
		cvsroot: /nfs/slac/g/jas/lcdroot
		userid: anoncvs
		password: jascvs
		module: lcd
Platform independent make (jmk) now used.
	Most development currently done on NT
	Now Unix development should be easy too


Set of experiment independent analysis tools for event oriented (High Energy Physics) data
	Data Access classes provide access to many common HEP data formats
	Histogram/Scatterplot Accumulation + Manipulation Classes
	Plot Display classes
	Lightweight framework for users to create physics analysis applications in Java.
Tools work alone, in combination, or within
	Java Analysis Studio GUI which gives:
		Integrated editor and compiler
		Efficient access to local and remote data
		Extensibility via Plug-ins, Fitters, Functions etc


	With many different simulated detectors and many physics processes, total MC data sample is large
	JAS has built in support for efficient distributed physics analysis


	LCD has set up central data repository at University of Pennsylvania (Don Benton), accessible from anywhere


Current (266Mhz PII, JDK 1.1.7)
		Clustering .6 secs/event
			13.5 Million Calorimeter Cells
		Fast MC 6 ms/event
		Track Finding + Fitting ~5secs/event
	Very competitive with C++/Root implementation (where they exist)
Will get even better!!!
		JDK 1.2, HotSpot - Run-time optimization
In real life may be faster than C++
		Better, cheaper performance analysis tools
		Java encourages lightweight, module interfaces which promote efficient coding styles
People time is what really matters


	Finish integration of MCFast Kalman filter
	Support for merging signal/backgrounds
	Additional Track Finders (projective, “cheater”)
	Improved Cluster Description
	Track/Cluster Association
	Cluster Refinement
	Vertex Finding code (based on ZVTop?)
	WIRED Event Display
	Support for SIO format reading/writing
	Switch to XML based geometry description
	Small Angle Tracking
	Reference Analyses


Come to the tutorial this afternoon
Works on Windows (95/98/NT] or Unix (Linux, Solaris,…) or Other
Online tutorial available
	Suitable for complete beginners:
		no knowledge of Java or JAS assumed
		starts with instructions on downloading and installing
		Shows simple sample analysis jobs
	http://www-sldnt.slac.stanford.edu/jas/documentation/lcd/
JAS Home Page
	http://www-sldnt.slac.stanford.edu/jas


	Many bug fixes - Printing Works - Save Histograms as GIF’s
	Easier to install LCD extensions
		no need to edit jas.ini
	Support for logging output to a file
	Supports 2D Binned Histograms (in addition to 1D Histograms and Scatter Plots)
	Standalone jobs can save histograms (for later viewing in JAS)
	Local jobs can append multiple datasets
	Built-in LCD documentation/help
	Gotcha’s for LCD users
		Must use JAS 20 version of lcd.jar (in lcd download area)
		Cannot connect to old servers (SLDNT0 OK)