Atlas:Analysis HC beyond STEP09

--Chollet 15:56, 19 octobre 2009 (CEST)

Distributed Analysis Stress Tests - HammerCloud beyond STEP09

Lessons learnt from STEP09

• Sites may identify reasonable amount of analysis they can assume and set hard limits on number of analysis running jobs
• Balancing data across many disk servers is essential.
• Very high i/o required by analysis (5 MB/s per job). Sites should review LAN architecture to avoid bottlenecks.

Results

• ATLAS HC Tests results : http://gangarobot.cern.ch/hc/ (New web interface!)
• ATLAS STEP09 summary : http://gangarobot.cern.ch/st/step09summary.html

ATLAS Info & Contacts

• Information via mailing list ATLAS-LCG-OP-L@in2p3.fr
• LPC : Nabil Ghodbane - Nabil.Ghodbane@cern.ch
• LAL : Nicolas Makovec
• LAPP : Stéphane Jézéquel
• CPPM : Emmanuel Le Guirriec
• LPSC : Sabine Crepe
• LPNHE : Tristan Beau
• CC-T2 : Catherine, Ghita
• IRFU : Nathalie Besson

HC Tests

Objectives

• Improve Cloud readiness by following site&ATLAS problems week by week (SL5 migration, site upgrades)
• Identify best data access method per site by comparing the event rate and CPU/Walltime

https://twiki.cern.ch/twiki/bin/view/Atla/HammerCloudDataAccess#FR_cloud

• Exercise Analysis with Conditions DB access (see where squid caching is needed) and Tag analysis

Data Access methods

Multiple data access methods are exercised

• via Panda : A copy-to-WN access mode using rfcp is used (xrootd in ANALY-LYON)
• via gLite WMS : 2 data access modes available
  • DQ2_LOCAL mode is a direct access mode using rfio or dcap
  • FILE_STAGER mode : data staged in by a dedicated thread running in // with Athena

Week 40

29/09/09 Test 649

• Muon Analysis (Release 15.3.1)
• Input DS (STEP09) : mc08.*merge.AOD.e*_s*_r6*tid*
• via Panda (mode copy-to-WN using ddcp/rfcp - xrootd in ANALY-LYON)
• http://gangarobot.cern.ch/hc/649/test/

  Bad efficiency - all sites affected all sites 
  Failed jobs with error : exit code 1137
  Put error: Error in copying the file from job workdir to localSE
  due to LFC ACL problem : write permissions in /grid/atlas/users/pathena
  for pilot jobs /atlas/Role=pilot and /atlas/fr/Role=pilot (newly activated)

30/09/09 Test 652, 653, 656, 657

• Muon Analysis (Release 15.3.1)
• Input DS (STEP09) : mc08.*merge.AOD.e*_s*_r6*tid*
• via WMS
• DQ2_LOCAL mode or direct access dcap/rfio : http://gangarobot.cern.ch/hc/652/test/
• DQ2_LOCAL mode or direct access dcap/rfio : http://gangarobot.cern.ch/hc/656/test/
• FILE_STAGER mode : http://gangarobot.cern.ch/hc/653/test/
• FILE_STAGER mode : http://gangarobot.cern.ch/hc/657/test/

http://lcg.in2p3.fr/wiki/images/ATLAS-HC300909.gif

Week 41

08/10/09 Test 663

• DPD Analysis (Release 15.5.0)
• Input DS - DATADISK : data09_cos.*.DPD*
• Cond DB access to Oracle in Lyon T1
• via Panda (mode copy-to-WN using ddcp/rfcp - xrootd in ANALY-LYON)
• Sites problems or downtime :
  • LAL : downtime
  • RO : DS unavailable
  • LYON (T2) : release 15.5.0 unavalaible

  Poor performance for foreign sites : Tokyo and Beijing compared to other french sites

http://lcg.in2p3.fr/wiki/images/HC663-081009-GRIF-Irfu-CPU.png http://lcg.in2p3.fr/wiki/images/HC663-081009-GRIF-Irfu-rate.png http://lcg.in2p3.fr/wiki/images/HC663-081009-Tokyo-CPU.png http://lcg.in2p3.fr/wiki/images/HC663-081009-Tokyo-rate.png

Recent talks

• ATLAS : from STEP09 towards first beams Graeme Stewart's talk@Journées Grille France (16 October 2009)
• Summary of HammerCloud Tests since STEP09 Dan van der Ster's talk@ATLAS Jamboree T1/T2/T3 (13 October 2009)
• HammerCloud Plans Johannes Elmsheuser's talk@ATLAS S&C Week (2 September 2009)