presentation

advertisement
The BaBar Mini
David N. Brown
Lawrence Berkeley National Lab
Representing the BaBar Collaboration
BaBar
BaBar’s Data Formats
Design of the Mini
Mini Performance and Status
The Mini in BaBar’s New Computing Model
BaBar
5-layer Si Vertex tracker
40-layer Drift Chamber
DIRC Cherenkov Counter
CsI Crystal Calorimeter
Muon Chambers in Fe
David N. Brown
LBNL
BaBar
2
CHEP03
25 March, 2003
BaBar Event Data Formats (2001)
Design Size
(Kbytes)
Actual Size
(Kbytes)
Online raw data
25
32
Reconstruction input
Objectivity raw data
25
50
Unused1
Reconstruction output
100
120
Unused1
Event Summary (ESD)
10
7
Unused1 2
Micro
1
3
Physics Analysis
0.1
1
Event Selection
?
4
Component navigation
Format
Event Tags
Event Header
1
2
Usage (2001)
The se formats were no t used excep t by a few experts
The E SD format was no t complete in 2001 , suppo rting on ly 3 of the 5 Babar sub systems
David N. Brown
LBNL
BaBar
3
CHEP03
25 March, 2003
A Gap in the Formats
 Raw, Reco, and ESD formats were not useful
 Reading any of these required staging many large files
 Time to read reco is comparable to running reconstruction
 A better persistent model was needed: the Mini
Transient
…
ESD
+
Reco
Kalman Fit
ESD
Reco Track
Kalman Fit
DC Hit
Persistent
Si Hit
…
Reco Track
Kalman Fit
DC Hit
Si Hit
Kalman Fit
Si Hit
Si Hit
Reco
Cluster
digi
Cluster
digi
digi
David N. Brown
LBNL
Cluster
Raw
digi
BaBar
4
digi
CHEP03
digi
Cluster
digi
25 March, 2003
digi
Mini Design Goals
 Support detector studies
 Provide low-level detector details to support common tasks
 calibration, alignment, diagnostics, and algorithm development
 Provide Reconstruction Object interfaces
 Support detailed Physics analysis
 Provide access to the full reconstruction results
 EG: track fits using Kaon mass for material effect predictions
 Allow users to follow calibration and alignment changes
 Allow detector-level systematic error checks
 Support a detailed event display
 Support the standard BaBar analysis interface
 Make it easy to access
 A disk size of 10KBytes/event or less
 A readback speed comparable to the Micro
 Allow customized output to fit specific needs
David N. Brown
LBNL
BaBar
5
CHEP03
25 March, 2003
Mini Design
 Directly persist high-level reconstruction objects
 Tracks, calorimeter clusters, PID results, …
 Indirectly persist lower-level reconstruction objects
 Track hits, calorimeter crystals, …
 Store ‘raw’ detector quantities (where possible)
 Digitization values, electronic channel id, …
 Pack data to detector precision
 Aggressively filter detector noise
 Avoid overhead in low-level ‘persistent’ classes
 Used fixed-size classes
 Align all data members
 No virtual functions in low-level classes
David N. Brown
LBNL
BaBar
6
CHEP03
25 March, 2003
Mini Persistence
 Pack data from low-level classes into compact objects
 Persist the entire transient tree in one persistent object
 References become indices into embedded arrays
 Every event fully described by 13 persistent objects
Transient
…
Reco Track
Kalman Fit
Persistent
Kalman Fit
Si Hit
Si Hit
Cluster
Cluster
DC Hit
digi
digi
David N. Brown
LBNL
digi
BaBar
digi
7
CHEP03
25 March, 2003
Data Packing
 Digitize floating point values
 Eg track fit parameters and covariance matrix
 Set packing precision at ~1% of detector resolution
 Use locally flat, globally logarithmic packing algorithm
 Packing precision depends on the value being packed
 Supports histograms without binning artifacts
 Bitwise OR small fields into packed data words
Track Impact parameter
Pack into 17 bits
longword alignment
Pack parameter error into
remaining 15 bits
David N. Brown
LBNL
BaBar
8
CHEP03
25 March, 2003
Mini Analysis Interface
 The BaBar Analysis Framework is entrenched
 Huge investment in Physicist code after 3 years of operation
 The original design supported multiple data formats
 But it had evolved to depend on details of the Micro
 Providing Mini-compatibility was a major effort
 Changes in the base classes, new subclasses, …
 The BaBar Mini Analysis interface is now working
 Physics (Micro) objects are built from native Mini objects
 Fully compatible with existing user code
 Provides access to most Mini-specific features
 Performance is comparable to reading the Micro
David N. Brown
LBNL
BaBar
9
CHEP03
25 March, 2003
Mini Performance
 Data size (after ootidy + gzip compression)
 6.5 Kbytes per generic physics event
 10 Kbytes per multi-hadron event (~10 tracks)
 Readback speed
 20 mSec per generic event (1GHz pentium III Linux)
Readback operation
% time
Transient creation + deletion
35
Objectivity data read
1030
Physics object creation
20
Framework overhead
10
Data field unpacking
0.1
David N. Brown
LBNL
BaBar
10
CHEP03
25 March, 2003
Single
SingleEvent
EventDisplayDisplay-Micro
Mini
David N. Brown
LBNL
BaBar
11
CHEP03
25 March, 2003
Mini Status
 The Mini was released for production in 2002
 All components of BaBar detector represented
 Trigger, MC truth matching, Particle ID, …
 ‘2002’ Reprocessing is nearly complete
 Should finish this week!
 Will provide the full Mini for BaBar’s full data sample
 Reco, Raw, and ESD were not written in this processing
 A Large savings in cpu, IO, tape, lock traffic, …
 Mini data is available at BaBar Tier-A sites
 The total Mini sample will be ~10 TeraBytes
 Access is through dynamic staging
 Small samples can be exported to smaller sites
 Physicists are starting to use it
David N. Brown
LBNL
BaBar
12
CHEP03
25 March, 2003
BaBar’s New Computing Model
 BaBar has recently revised its computing model
 Prompted (partly) by the need to integrate the Mini
 The Mini will be ported to use RootIO
 Allows interactive (CINT) access to production output
 Embedded arrays will be converted to Root columns
 Embedded object classes will be directly reused
 A Reduced Mini will replace the existing Micro
 Cluster Mini objects used directly in analysis into New Micro
 Cluster other objects separately to complete the Mini
 The Analysis interface will be re-implemented
 Optimized for Mini access
David N. Brown
LBNL
BaBar
13
CHEP03
25 March, 2003
Conclusions
 BaBar has implemented a new Event format: the Mini
 Replaces inefficient Raw, Reco, and ESD formats
 Provides access to detector detail for average users
 The full BaBar data sample will soon be available in
Mini format
 BaBar is implementing a new Computing Model
 The Mini will be ported to RootIO
 The Micro will be replaced with a Reduced Mini
 The Analysis interface will be re-implemented
 The new model will be deployed in late 2003
 We are close to achieving our original goal of a
flexible, unified, efficient event data format
David N. Brown
LBNL
BaBar
14
CHEP03
25 March, 2003
Download