ATLAS Computing Model

1
ATLAS Computing Model

• LCG Resources Group
• CERN, 26th March 2004
• Roger Jones, ICB Chair

2
Inputs

• ATLAS internal
• FORTRAN/ZEBRA timings and event sizes
• Current OO/C timings and sizes
• Relevant expert inputs
• External
• LEP experience and analysis patterns
• D0 computing model
• Beginning to look at CDF computing model
• Data Challenges
• Test the computing model
• Refine the code and timings
• Efficiencies are not factored in (movable feast)

3
Nominal year 107 s Accelerator efficiency
50
4
Processing times

• Reconstruction
• Time/event for Reconstruction now 60 kSI2k sec
  
• We think we can reasonably recover a factor 4 at
  least
• factor 2 from running only one default algorithm
• factor 2 from optimization
• Therefore now we take as reference 15 kSI2k
  sec/event
• Simulation
• Time/event for Simulation now 400 kSI2k sec
• This is the worst case right now.
• For resource calculations, let us assume
• factor 2 from optimization (work already in
  progress)
• factor 2 on average from the mixture of different
  physics processes (and rapidity ranges)
• Therefore now we take as reference 100 kSI2k
  sec/event
• Number of simulated events needed 108
  events/year (??)
• Note D0 simulate 10 of real data rate as here
• Generate samples about 3-6 times the size of
  their streamed AOD samples

5
Operation of Tier-0

• The Tier-0 facility at CERN will have to
• Hold a copy of all raw data to tape
• Copy in real time all raw data to Tier-1s
  (second copy useful also for later reprocessing)
• Keep calibration data on disk
• Run first-pass calibration/alignment and
  reconstruction
• Distribute ESDs to external Tier-1s
• (1/3 to each one of 6 Tier-1s)
• Currently under discussion
• Archiving of simulated data probably not
• Sharing of facilities between HLT and Tier-0
• reprocessing capacity versus architecture
• Distribution and access to conditions data

6
Operation of Tier-1s and Tier-2s

• We envisage at least 6 Tier-1s for ATLAS. Each
  one will
• Keep on disk 1/3 of the ESDs and a full AODs
  and TAGs
• Keep on tape 1/6 of Raw Data
• Keep on disk 1/3 of currently simulated ESDs and
  on tape 1/6 of previous versions
• Provide facilities for physics group controlled
  ESD analysis
• Calibration and/or reprocessing of real data (one
  per year)
• We estimate 4 Tier-2s (various sizes) for each
  Tier-1. Each one will
• Keep on disk a full copy of TAG and roughly one
  full AOD copy per four T2s
• Keep on disk a small selected sample of ESDs
• Provide facilities (CPU and disk space) for user
  analysis and user simulation (25 users/Tier-2)
• Run central simulation

7
Analysis model (1)

• Analysis model broken into two components
• Central production of tuples and TAG collections
  from ESD
• User analysis of group tuples/streams, new
  selections etc and total user simulation matching
  the official production
• Central analysis
• Assume 20 analysis WGs
• Each analysis channel may require several
  iterations of event selections, starting with all
  TAGs, then with less AOD, much less ESD and very
  few RAW events
• physics working groups will report on analysis
  model later this year
• experience from DC2 exercise will provide useful
  information
• 100 passes over ESD in total, producing tuples
• Retain current and previous tuple on disk
• Estimate data reduction to 10 of full AOD, so
  720Gb/group/annum
• 0.5kSI2k per event (estimate), quasi real time ?
  9MSI2k
• This is 3 times the first pass reconstruction load

8
Analysis model (2)

• User analysis
• Each user (600 total) assumed to
• Perform 1/N of the MC non-central simulation load
• Perform analysis of WG samples and AOD
• Analyse 10 of WG statistics (1 total), 25
  passes/year
• Retain current and previous tuple on disk ?
  2x0.9Tb
• 0.5kSI2k per event, full year to process ?
  4.2kSI2k
• Perform private simulation, 1/600th of total MC
  sample
• 0.6kSI2k required, 435Gb on disk and tape
• Total requirement 4.7kSI2k and 1.5/1.5Tb
  disk/tape
• Assume this is all done on T2s

9
CERN Tier-2/1

• In model, CERN non-T0 primarily for large user
  community
• Tier-0 and a large Tier-2?
• Could do some reprocessing
• but just use staging pool and T0 tape, no extra
  copies
• but external T1s can cope
• Do not assume simulation role
• 100 ATLAS users assumed, 4xnormal T2

10
CERN T1/2 in this Model
472 kSI2k required
11
External Tier 1
Processing for Physics Groups 1500
kSI2k Reconstruction 508 kSI2k
12
A Regular T2
Simulation 17 kSI2k Reconstruction 2 kSI2k Users
118 kSI2k Total 137 kSI2k
13
Summary (1)
Without Efficiencies
14
Summary (2)
Efficiencies (added during meeting) Scheduled cpu
85, chaotic 60 Disk 70