ATLAS calibration and alignment strategy by wbf21256


									                     ATLAS calibration and alignment strategy

                                 Richard Hawkings
                            ATLAS plenary meeting, 18/2/05

 Steps in ATLAS calibration / alignment
 Subdetector calibration questionnaire and responses
       Calibration in ROD, HLT, Tier-0 and offline
       Use of dedicated calibration streams
       Latency and remote calibration issues
       Offline calibration
 Conclusions and open issues

     Many thanks to subdetectors for useful discussions, and Fabiola Gianotti for
      collaboration on questionnaire and associated writeup

18th February 2005                    Richard Hawkings                               1
                          Calibration steps and strategies

 Overview of calibration (+alignment) evolution:
     Calibration preparation before detectors are installed
          Work done in institutes and at CERN – results mainly in ‘subdetector’ databases
     Calibration during commissioning
          In-situ electronics calibration, cosmics, single beam running
          Defines the calibration for day 1 – using ATLAS conditions database by this point
     Calibration from physics data
          Calibration determined in RODs and HLT
          Dedicated calibration step before prompt reconstruction
          Continuing ‘offline’ calibration procedures to refine constants for subsequent data
           reprocessing – process will continue for months and years after data is taken.
 Calibration consumers
     Online/trigger system (need fast and reliable calibration), Tier-0 prompt
      reconstruction, later reconstruction passes (primarily at Tier-1 centres)
   What are the subdetectors’ plans?
     Little systematic knowledge at ATLAS level
          Are they compatible / consistent ?
     Need to know for planning computing model (online/offline resources), DC3
      ‘calibration/alignment closed loop test’ and for ongoing commissioning activities
18th February 2005                       Richard Hawkings                                        2
                          Questionnaire and responses

 Subdetector calibration/alignment questionnaire exercise in Dec 04/Jan 05
 Main questions:
     Where in processing chain will calibrations be performed?
          What are the associated CPU power requirements?
     What dedicated calibration streams are required as part of event filter output?
     Need for a ‘prompt calibration’ step between event filter and prompt
     Need to stream calibration events from event filter to remote institutions?
     What are the offline processing requirements (for ‘final’ calibrations) – samples
      needed and access to RAW and ESD data?
     How will requirements evolve between start-up and steady-state running?
 All subdetectors have now responded
     Level of detail and advancement in understanding/planning varies a lot …
          Active discussions going on in many subdetector communities
     Summary of responses in main areas… for more details see note at

18th February 2005                     Richard Hawkings                                   3
                                Calibration at ROD level

 ROD-level calibration: performed outside physics and during physics runs
     Data generally processed inside RODs (no event building), summary information
      (histograms, generated calibrations) passed to offline and database
          Larger data readout requests for initial debugging
     Several requests for between-run/daily calibration tasks of  1 hour duration
          SCT/pixel electroncis calibration, LAr ramp runs, tile laser/charge injection/pedestal,
           RPC and CSC pulsers
     Some longer tasks of several hours – 1 day duration, expect to perform ~monthly
          LAr delay runs, Tilecal cesium source calibration
          Level 1 trigger calibration (cooperation of e.g. calorimeter and level 1) – especially
           during initial startup with beam
     ROD level calibration/monitoring during physics
          Many detectors – dead/noisy channels, pedestals, t0 monitoring, efficiencies
     CPU requirements generally modest (subdetector workstations?), anticipate use
      of some event filter resources for partial event building outside physics

18th February 2005                        Richard Hawkings                                           4
                                 Calibration in the HLT

 Calibration tasks running in HLT:
 Output of data to dedicated calibration streams, little CPU-intensive work
     ID has dedicated stream of high pT tracks for alignment and TRT calibration
          Processed track, hit and residual data – specialised data format for alignment
     LAr wants to perform (part of) Zee calibration in event filter
     Muon system outputs dedicated stream of O(1 kHz) precision/trigger hits for t0,
      autocalibration and alignment
          Hit information in very restricted road + initial muon trackfit
          Most promising avenue is to extract information from LVL2 muon trigger processing
          Ongoing work to check implications (CPU, event collection) – not originally foreseen
     Level-1 trigger will use calorimeter pulsers to check calo/trigger gains/responses
 Lots of monitoring will happen at this stage – not explicitly asked about in
     Good place to perform generic data-quality monitoring on HLT-accepted events
     Need to understand CPU requirements on event filter farm – monitoring is a
      secondary task of HLT system

18th February 2005                       Richard Hawkings                                         5
                       Calibration streams from event filter

 Identifying a detailed list of calibration streams requested from event filter
 Streams with partial readout of single/multiple detectors, restricted ROI
      ID generic high-pT tracks for alignment (10-100 Hz, specialised format)
      LAr readout of 5-sample calorimeter RAW data in ROI around high-pT electrons
           50 Hz, possibly phase out after few months of running
      High pT-muons identified at LVL1, processed through LVL2 (1 kHz, specialised
       event format)
      High pT-muons in large/small chamber overlap regions for alignment (~5 Hz)
      Isolated high-pT hadrons (from single-prong  trigger?) for calorimeters and TRT
 Streams with full event readout – duplicate events in physics stream
      Inclusive e/ with e.g., pT>20 GeV, dileptons and prescaled minimum bias
      Duplicate and separate these events for fast efficient access for detector
       calibration experts – especially important during first data-taking
 Streams sum to 40-50 MB/sec, i.e. ~15% of datataking bandwidth
      Many events identified late on in EF processing – have to be collected from all
       event filter output nodes – data collection and bookkeeping issues
 18th February 2005                      Richard Hawkings                                6
                       Processing requirements at Tier 0

 ATLAS computing model allocates 500 kSI2k units (100 dual 8GHz CPUs)
  to calibration activities at Tier-0 (~13% total capacity in 2008)
 Identified subdetector requests for ‘Tier-0’ CPU capacity to process special
  calibration streams in preparation for prompt reconstruction
       SCT and pixel alignment: 50 kSI2k
       TRT alignment and calibration: 20 kSI2k
       MDT t0 and autocalibration: 130 kSI2k
       TGC alignment, calibration and efficiency determination: 50 kSi2k
       RPC level 1 trigger calibration: 10 kSI2k
 Estimates are very preliminary – nothing from calorimeters yet
     Identified requests total 260 kSI2k units (but no calorimeters yet), cf. computing
      model allocation of 500 kSI2k
          Reasonable match – most Tier-0 calibration resources will be devoted to prompt
           calibration to allow prompt reconstruction to proceed.
          Also need to assess disk space requirements (240 TB allocated in computing model)

18th February 2005                     Richard Hawkings                                    7
                          Latency and remote calibration

 Latency – how long between end of fill and start of prompt reconstruction?
     Computing model document proposed 24 hours – time to process calibration
      stream, generate calibration and do some verification (including manual check)
          NB: Separate ‘express line’ forseen for faster processing of ‘discovery-type’ events,
           without waiting for calibration iteration
     All subdetectors feel 24 hours is enough (providing processing power is available)
          Also natural timescale for combining asynchronous calibration (e.g. optical alignment)
     Longer latency can be expected at initial startup
          Less-automated procedures, process samples over and over again
          Will want to process bulk physics sample multiple times
 Remote calibration
     No specific requests for streaming calibration data to remote institutes
          But recognition that this might be useful, especially if Tier-0 resources or computing
           infrastructure to access CERN are insufficient
          Plan to do ‘prompt’ calibration at CERN, but keep possibility open (also for monitoring)?
     Offline calibration (after prompt reconstruction) will be much more geographically
      distributed (as for ‘analysis’ tasks) – involving the whole collaboration
          Data distribution and network implications yet to be assessed in detail

18th February 2005                       Richard Hawkings                                          8
                                    Offline calibration

 Prompt calibration aims at providing constants for first-pass reconstruction
     Subsequent offline calibration steps are needed to refine calibration/alignment to
      extract ultimate performance of ATLAS
          Need more statistics, studies over long time periods
 Generally uses ‘well-known’ physics channels, e.g:
     Inclusive electrons and muons (>20 GeV pT ?)
     Z,J/, decays to lepton pairs (+ radiative photons) and W
     /Z + jet and multijet, tt events, …
 Need to understand data access patterns, especially for RAW data
     Now start to see first definitions of ESD and AOD data – what can be done?
          Can ESD contents be improved to reduce need to access RAW?
     Samples can be accessed from calibration, express and physics streams – what
      will be most efficient for ‘long-term’ processing?
     Most offline calibration activities will be based around Tier 1/ Tier 2 centres
     Need to bring calibration constants back to central conditions DB at CERN
     In general, expect 2-3 months between prompt and first re-processing
18th February 2005                       Richard Hawkings                             9
                            Conclusions and open issues

 Very useful first exercise to understand subdetector calibration requirements
     Resource needs: some CPU for ROD processing, and substantial Tier-0 CPU
     Little requirement for calibration CPU in HLT, but remember monitoring!
 Identified issues to be followed up:
     Calibration streams require partial event building (by detector, by ROI) dependent
      on event type – implications for TDAQ dataflow?
          How do we collect and catalogue small calibration streams?
     Writing of substantial data quantities directly from RODs, perhaps with several
      detectors operating together (e.g. LVL1+calorimeters)
     Requests to write more data at startup (e.g. LAr 5 samples) – bandwidth vs trigger
      rate considerations
     Calibration stream selections need detailed studies
          Thresholds, rates and purities with ‘as built’ detector calibration
          Single isolated hadron sample needs particular study
 Fast efficient calibration will be key at ATLAS startup
     Start to exercise calibration plans in commissioning and DC3

18th February 2005                         Richard Hawkings                         10

To top