110 likes | 181 Views
This report by Ilija Vukotic discusses MC event sizes, streams sizes, categories, IO performance, and possibilities for improvement in ATLAS Software. It includes detailed information on data sizes, event sizes, and recommendations for enhancements in performance and efficiency.
E N D
Data size and IO performance report ATLAS Software & Computing Workshop
Data size • MC http://athena-infoioperformance.web.cern.ch/athena-infoIOperformance/MCsizes.html • Streams http://athena-infoioperformance.web.cern.ch/athena-infoIOperformance/StreamsSizes.html • Categories http://athena-infoioperformance.web.cern.ch/athena-infoIOperformance/T0streams.html • IO performance • Possibilities for improvement Ilija Vukotic
MC event size Ilija Vukotic
MC event size *Compared to 22 Aug. 2010 Note: trigger size is not realistic as used trigger menu was not realistic. Problem rectified with addition of new aliases: Physics_default -> Physics_pp_v1 MC_loose_default -> MC_pp_v1_loose_mc_prescale MC_tight_default -> MC_pp_v1_tight_Mc_prescale (see bug report #74712). Ilija Vukotic
Streams sizes • Only runs with 5k+ events • and good LBs • Shows: • AODs,ESDs and all DESDs • Events/total size/ev. size Ilija Vukotic
Streams sizes ESD DESDM_EGAMMA AOD DESD_SGLMU DESD_SGLEL Ilija Vukotic
Categories Ilija Vukotic
Categories JetTauETMiss Muons L1Calo Egamma Details and list of collections in categories can be found at: https://spreadsheets.google.com/ccc?key=0AiPvgbRljNCodG5IY1BseHRHa2R5WjZoMGRnLXp0dHc&hl=en Ilija Vukotic
IO performance Slow Bad CF KEEP IN MIND In real data TrackCollections are even larger ! Ilija Vukotic
IO performance • Do we really need all of these? • No way to say if object is ever used. • What is so terrible about having a name attached to each object we store? Ilija Vukotic
Possibilities for improvement • TrackCollection/TrackParticle container • Maybe 50kb in size and factor 2 in time • Few MB in memory • CaloShowerContainer • Probably can go down in size • Calo(Topo)TowerContainer • Gets recreated at read time • Must be possible to optimize further • TRT_Drift_Circle_Container • Must be faster and smaller My estimate: In total between 50 and 200kB/ev can be shaved in converters Should be possible >100ms/ev (currently at 480 ms/ev) Other options: Object removal. Cuts tuning Ilija Vukotic