gridpp13 atlas

ATLAS and Grid Computing : ATLAS and Grid Computing RWL Jones GridPP 13 5th July 2005


ATLAS Computing Timeline : ATLAS Computing Timeline Commissioning takes priority!


Computing TDR structure : Computing TDR structure The TDR describes the whole Software & Computing Project as defined within the ATLAS organization:


Massive productions on 3 Grids : Massive productions on 3 Grids


Massive productions on 3 Grids (3) : Massive productions on 3 Grids (3) July-September 2004: DC2 Geant-4 simulation (long jobs) 40% on LCG/EGEE Grid, 30% on Grid3 and 30% on NorduGrid February-May 2005: Rome production 70% on LCG/EGEE Grid, 25% on Grid3, 5% on NorduGrid LCG/EGEE Grid resources always difficult to saturate with “traditional” means New approach (Lexor-CondorG) used Condor-G to submit directly to the sites in this way the job rate was doubled on the same total available resources much more efficient usage of the CPU resources the same approach is now evaluated also for the Grid3/OSG Grid job submission which suffered also from job rate problems


Massive productions on 3 Grids (4) : Massive productions on 3 Grids (4) 73 data sets containing 6.1M events simulated and reconstructed (without pile-up) Total simulated data: 8.5M events Pile-up done later (for 1.3M events done up to last week)


Experience with LCG-2 Operations : Experience with LCG-2 Operations Support for our productions was excellent from the CERN-IT-EIS team Other LCG/EGEE structures were effectively invisible (GOC, ROCs, GGUS etc) no communication line between experiments and the Grid Operations Centres operational trouble info always through the EIS group sites scheduled major upgrades or downtimes during our productions no concept of “service” for the service providers yet! many sites consider themselves as part of a test structure set up (and funded) by EGEE but we consider the LCG Grid as an operational service for us! many sites do not have the concept of “permanent disk storage” in a Storage Element if they change something in their filing system, our catalogue has to be updated!


Second ProdSys development cycle : Second ProdSys development cycle The experience with DC2 and the Rome production taught us that we had to re-think at least some of the ProdSys components The ProdSys review defined the way forward: Frederic Brochu one of the reviewers Keep the global ProdSys architecture (system decomposition) Replace or re-work all individual components to address the identified shortcomings of Grid middleware: reliability and fault tolerance first of all Re-design the Distributed Data Management system to avoid single points of failure and scaling problems Work is now underway target is end of Summer for integration tests ready for LCG Service Challenge 3 from October onwards


Distributed Data Management : Distributed Data Management Accessing distributed data on the Grid is not a simple task Several central DBs are needed to hold dataset information “Local” catalogues hold information on local data storage The new DDM system (right) is under test this summer It will be used for all ATLAS data from October on (LCG Service Challenge 3) Affects GridPP effort


Computing Operations : Computing Operations The Computing Operations organization likely to change: Grid Tools Grid operations: Tier-0 operations re-processing of real and simulated data at Tier-1's data distribution and placement Software distribution and installation Site and software installation validation and monitoring Coordination of Service Challenges in 2005-2006 User Support Proposal to use Frederic Brochu in front-line triage Credited contribution Contingent on Distributed Analysis planning


Software Installation : Software Installation Software installation continues to be a challenge Rapid roll-out of release to the Grid important for ATLAS UK eScience goals (3.1.4) Vital for user code in distributed analysis Grigori Rybkine (50/50 GridPP/ATLAS eScience): Working towards 3.1.5, kit installation and package management in distributed analysis Package manager implementation supports tarball and locally-built code Essential support role 3.1.5 progressing well, 3.1.4 may have some delays because of external effort in nightly deployable packages


Current plans for EGEE/gLite : Current plans for EGEE/gLite Ready to test new components as soon as they are released from the internal certification process assume the LCG Baseline Services Only seen the File Transfer Service & LCG File Catalogue both being actively tested by our DDM group FTS will be field-tested by Service Challenge 3 starting in July LFC is in our plan for the new DDM (Summer deployment) Not really seen the new Workload Management System nor the new Computing Element some ATLAS informal access to pre-release versions As soon as the performance is acceptable we will ask to have them deployed this is NOT a blank check!


Distributed Analysis System : Distributed Analysis System ATLAS and GANGA work now focused on Distributed Analysis LCG RTAG 11 in 2003 did not produce a common analysis system project as hoped. ATLAS therefore planned to combine the strengths of various existing prototypes: GANGA provides a Grid front-end for Gaudi/Athena jobs DIAL provides fast, quasi-interactive, access to large local clusters The ATLAS Production System to interface to the 3 Grid flavours Alvin Tan Work on the job-building GUI and Job Options Editor well received Wish from LBL to merge JOE with Job Options Tracer project Monitoring work also well received – prototypes perform well. Frederic Brochu Provided beta version of new job submission from GANGA direct to Production System


Distributed Analysis System (2) : Distributed Analysis System (2) Currently reviewing this activity to define a baseline for the development of start-up Distributed Analysis System All this has to work together with the DDM system described earlier Decide a baseline “now”, so we can have a testable system by this autumn The outcome of the review may change GridPP plans


Conclusions : Conclusions ATLAS is (finally) getting effective throughput from LCG The UK effort is making an important contribution The Distributed Analysis is continuing to pose a big challenge ATLAS is taking the right management approach GridPP effort will have to be responsive