The PHysics Analysis SERver Project (PHASER)

1
The PHysics Analysis SERver Project(PHASER)
M. Bowen, G. Landsberg, and R. Partridge Brown
University

• CHEP 2000
• Padova, Italy
• February 7-11, 2000

2
What is the PHASER project?

• Effort to substantially increase productivity of
  physicists analyzing multi-TB summary data sets
• Our immediate focus is on the DØ experiment
• 600 million data events/year starting in early
  2001
• Summary data set expected to grow at rate of
  3TB/year
• Concentrate on event selection and ntuple
  creation stage
• transition in data handling from monolithic
  reconstruction processing to the much more
  chaotic processing of summary data by many
  physicisits
• IO and CPU intensive due to need to apply latest
  calibration, particle ID, and event selection
  algorithms to several hundred million events

3
PHASER Architecture

• Physics Object Database (POD) stores meta-data
  used by most physics analyses for their initial
  event selection
• Physics Object and Particle ID tables in POD
  store calibrated 4-vectors, object quality
  variables, and results of particle ID algorithms
• DVD storage of full summary (mDST) data set and
  useful subsets of larger DST and STA data sets

4
PHASER is PHast

• New calibrations and particle ID algorithms can
  be quickly incorporated
• Only the changes need to be importd
• Regenerating the large mDST data set will only be
  done infrequently
• Storage of up-to-date calibrations and particle
  ID algorihtms avoids the need to re-apply these
  alogorithms for each event selection pass
• Particle ID tables are small, making it possible
  to quickly eliminate events not having the
  desired set of physics objects
• Direct access to full mDST sample on DVD allows a
  mDST subset to be quickly generated for advanced
  analyses developing new algorithms not yet in the
  database

5
The Physics Object Database (POD)

• Stores fully calibrated meta-data associated with
  the various physics objects
• leptons, photons, jets, missing ET, secondary
  vertices, triggers, etc.
• for example, an electron object would have the
  energy, direction, and various quantities used in
  the electron ID algorithms stored
• Each physics object associated with a table in a
  relational database
• Primary key uniquely identifies each physics
  object and provides information needed to
  correlate physics objects from a single event
• Currently use Run, Event, Instance (where
  appropriate) and row number from ntuple used to
  load database
• Alternative data source index, sequence number,
  and instance

6
Why use a Relational Database?

• Physics objects typically have a fixed set of
  attributes used for event selection and analysis
• Independence of tables aids loading, updating
  database
• Data can be bulk loaded as long as primary key
  is provided in input data stream
• Several vendors with quite capable products,
  large commercial market

7
Prototype POD

• Use DØ Run 1 data (1992 - 1996 running period)
• 62 million events loaded into the database
• Entire All-Stream data set loaded
• Data set used by almost all DØ physics analyses
• Only files with special processing or trigger
  conditions excluded
• Column-wise ntuple format used for
  importing/exporting data

8
DØ Run 1 POD

• Including indexes, Run 1 POD occupies 100 GB
• 58 physics object data
• 18 indexes on object ET
• 12 primary keys
• 12 database overhead

9
POD Benchmarks

• Z ? ee- candidate event selection
• 7 seconds to identify 6k events
• W ? en candidate event selection
• 18 seconds to identify 86k events
• Both benchmarks times make use of particle ID
  tables
• Event selection times compare very favorably with
  1000 CPU hours required to generate ntuples used
  in this study
• Benchmark Hardware/Software
• 450 MHz dual-processor Pentium II with 256 MB RAM
• Database stored on (6) 36 GB disks in Raid 0
  stripe set
• MS SQL Server running on Windows NT 4.0

10
DVD Storage

• Provide access to additional event information
  not included in POD
• DVD-RAM has a number of unique capabilities
• Less expensive than disk storage, doesnt require
  backup
• Access to individual events is much faster than
  tape storage
• Current disk capacity is 2.6 GB, 4.7 GB expected
  soon
• Commercial DVD libraries hold up to 600 DVD disks
• 2.8 TB capacity using 4.7 GB DVD-RAM disks
• Average disk load time of 4.5 s, lt1 hour to cycle
  through 600 disks
• Up to 6 DVD-RAM drives gives 10 MB/s IO rate

11
Web Interface

• Plan to develop web-based user interface
• Interface modelled on 3-tier architecture
  widely used in commercial applications
• Physicist will enter event selection requirements
  using a Java applet
• Applet communicates request to Physics
  Intelligence middleware running on PHASER system
  (via CORBA)
• Translate request to SQL for event selection
• Verify that request can be accommodated within
  resource constraints
• Produce the requested output files

12
PHASER Output

• Several output options
• List of run and event numbers satisfying the
  request
• Ntuple created from POD information
• mDST stream containing requested events from DVD
  library
• Output files will generally be small enough to
  transfer over the network
• Larger output files can be written to DVD and
  physically sent to physicist for further analysis

13
Conclusions

• PHASER offers a way for both experts, novices,
  and dinosaurs to quickly extract information
  about a particular class of events
• Feasibility of loading Run 1 size physics
  object info into a relational database has been
  demonstrated
• Significant improvements in event selection time
  has been observed for W/Z benchmarks
• Expect these results will scale up to Run 2 data
  load
• Database technology is also potentially useful
  for helping manage complex analyses and storing
  intermediate results