Upload
edmund-young
View
228
Download
0
Tags:
Embed Size (px)
Citation preview
Horst Fischer University Freiburg /
CERN
COMPASS Computing
Outline
•Central Data Recording and Event Database
•Event Reconstruction
•Some words on statistics
•Analysis Model
A word of warning:2001 COMPASS Detector commissioning2002 COMPASS Analysis commissioningHence, some numbers are preliminary and not yet on rock solid ground
Long-distanceG igabit E thernet
C P U C P U C P UC P UC P U C P UC P U C P U C P UC P UC P U C P U
C P U C P U C P UC P UC P U C P UC P U C P U C P UC P UC P U C P U
C P U C P U C P UC P UC P U C P UC P U C P U C P UC P UC P U C P U
C P U C P U C P UC P UC P U C P UC P U C P U C P UC P UC P U C P U
C P U C P U C P UC P UC P U C P UC P U C P U C P UC P UC P U C P U
C P U C P U C P UC P UC P U C P UC P U C P U C P UC P UC P U C P U
C P U C P U C P UC P UC P U C P UC P U C P U C P UC P UC P U C P U
C P U C P U C P UC P UC P U C P UC P U C P U C P UC P UC P U C P U
C ASTO RD ata Server
D ata Server
D ata Server
D ata Server
D ata Server
D ata Server
35 MB/s input rate
Central Data RecordingCentral Data Recording
Data servers:Up to 20 x 500GB During shutdown: Use disk of online farmas additional stage pool(additional 3.5 TB)
CASTOR:CERN development of a“infinite file system”COMPASS is the first experiment which uses CASTOR heavily
DST production
tape
Central Data Recording
Design value (35MB/s)
260 TByte in ~100 days
Compare to BaBar:
1TByte/day
662 TByte in 1999-2002
Or CDF: 500 TByte/year
(U.S. year >> 100 days)
May 27 Sep 18
GB
GB
Data in Database
2001:Detector commissioning: 65 days, 13 TByteTarget polarized longitudinally: 14 days, 15 TByte total: 28 TByte
2002:Target polarized longitudinally: 57 days, 173k spillsTarget polarized transversely: 19 days, 52k spills total: 260 TByte
2002 on average 22k triggers/spill 5 Gev (3.8Gev, 1.2Gev) online: ~260,000 files (~80 files /
run)
Objectivity faced out at CERN and replaced by Oracle9i
After transfer to CCF ~2000ev/s are formatted into Objectivity/DB Different data bases for raw events and special events Planed: event reconstruction before storing DB federations to tape 2002: offline reconstruction (calibrations were/are not available)
Database Migration
Raw events• Original DATE format, flat files kept in Castor
Metadata• Mainly event headers and navigational information for raw and
reconstructed events in relational database• 0.25% of the raw data volume• at CERN stored in Oracle (but without Oracle-specific features)• Possibility for another database in the outside institutes
(MySQL)
Conditions data• Migrated to the new CDB implementation based on Oracle• No interface change (abstract interface)
DST• Object streaming to files
More on DB Migration
CERN IT/DB is involved in the migration project to a large extent
Planning activities started early this yearSeveral people (IT/DB/PDM section) working on the projectCoordinating experiments and with other IT groups
– Servers: ADC group
– Castor: DS groupMigrate database to “borrowed” tapes
COMPASSIT/DB is working closely with COMPASS expertsCertify interfaces to Oracle DB
Licenses being negotiated between CERN and Oracle for >2003
Event Reconstruction
Compass Computing Farm: 100 dual Processor PIII Linux PCs Mainly for event reconstruction and (Mini)-DST
production maintained and operated by CERN staff Part of CERN LSF BATCH-system
Average time to reconstruct one event: 400ms/ev (300…500 ms, during the first 400ms of a spill up to 1300ms/ev)
5Gev * 400ms/ev = 2Gs / 200CPUs = 10 Ms/CPU at 100% efficiency: = 115 days on 200
CPUs
Today: about 30% of 2002 data are being pre-processedfor preliminary physics analysis (since September)
Process as much in parallel as possible: 1 run : ~ 80 data files
: ~ 80 batch jobs running at the same time
Batch Statistics
Jobs per weekIt’s a full time job to control data flow and quality of output
Must set-up
production shifts
to share work load
like one does for
data taking
Alternative: Outsourcing
= hire technicians?
CPU time per week
COMPASS Analysis Model
CalibrationsMySQL
DSTOracle 9i
mini DSTROOT tree
RAW dataOracle 9i
CORAL
PHAST
micro DSTROOT tree
micro DSTROOT treemicro DST
ROOT tree
ConditionsPVSS/Oracle 9i
e-LogbookMySQL
CORAL - Event Reconstruction
= COMPASS Reconstruction
and AnaLysis Program Modular architecture
Following OO techniques
Fully written in C++
Written from scratch
Defined interfaces for easy exchange of external packages (e.g. OB/DB Oracle9i
Access to event, conditions and calibration data bases
Not as user friendly, however, as it may look
Calibrations etc.
Presently (2001 & 2002):Presently (2001 & 2002):
Calibration DB: 42 MByte
Map DB: 4.5 MByte
Geometry DB: 4.4 MByte
Calibrations stored in Run dependent file structured data base which allows for different versions Bookkeeping: MySQL (project launched recently)Calibrations: Alignment (Torino)
RICH refraction index, PID likelihood (Trieste)
Time-zero calibrations (Bonn, Erlangen & Mainz)
r-t relations (Saclay & Munich) Amplitude and time for GEM & Silicon
(Munich) Calorimeter (Dubna & Protvino)
Detector calibrations can not be automated yet. Still everyday new surprises …
Is a drain for our most valuable resource “manpower”Tasks must stay in responsibility of the listed groups, otherwise very inefficient because of learning curve
Experience: Reconstruction & Analysis
(my personal, subjective, view)
“Distributed” data analysis is very inefficient (presently!)Size of individual groups participating in analysis is not yet large enough to make fast & significant progress by their own
•Frequently struggling with infrastructural problems•Communication between satellites is poor or not existing•Missing discussions & stimulations This unsatisfactory situation may improve with time (?)
COMPASS needs a strong & centralized analysis group
Exchange with satellite groups due to travel of people coming/leaving from/to home institutes
Short term SOLUTION:
Analysis: Where & Resources
significant core of experts @
CERN
“Analysis Center”
Resources: CCF+…
Trieste(55 GHz, 4 TB)
TU Munich(26 GHz, 2.4TB)2003: double
GridKa (German LHC Tier 1)
530 GHz, 50TBCOMPASS: 20 GHz
2003: double
Freiburg(40 GHz, 3TB)
2003: *1.5
Dubna(14.5 GHz,1.7TB)
2003: double
Saclay(Lyon)
Mainz/Bonn
Torino(66 GHz, 1 TB)
Conclusions & Suggestions
COMPASS faces its first year of “real” data analysisChallenges are enormous: calibrations, data handling, etc.… but very interesting physics ahead and thus its worth all the efforts (and sleepless nights)
Unfortunately we are forced to change event data base in a moment when we should work with full pace on our first data analysis (disadvantage when you are dependent on third party decisions!)
Tools for event reconstruction and (Mini- & Micro-) DST analysis exist but need continuous improvements and adaptations to reality
Still weak on tools to submit & monitor and manipulate & handle output
of 200+ batch jobs per hour (effort was underestimated by many of us)
Presently under evaluation: ALIEN (ALICE development)
Organizational structure of analysis group: 1. Establish a strong core group at CERN with regular
WORK(!)shops2. Setup shifts for DST production