Job optimization in ATLAS TAG-based distributed analysis

The ATLAS experiment is projected to collect over one billion events/year during the first few years of operation. The efficient selection of events for various physics analyses across all appropriate samples presents a significant technical challenge. ATLAS computing infrastructure leverages the Grid to tackle the analysis across large samples by organizing data into a hierarchical structure and exploiting distributed computing to churn through the computations. This includes events at different stages of processing: RAW, ESD (Event Summary Data), AOD (Analysis Object Data), DPD (Derived Physics Data). Event Level Metadata Tags (TAGs) contain information about each event stored using multiple technologies accessible by POOL and various web services. This allows users to apply selection cuts on quantities of interest across the entire sample to compile a subset of events that are appropriate for their analysis. This paper describes new methods for organizing jobs using the TAGs criteria to analyze ATLAS data. It further compares different access patterns to the event data and explores ways to partition the workload for event selection and analysis. Here analysis is defined as a broader set of event processing tasks including event selection and reduction operations ("skimming", "slimming" and "thinning") as well as DPD making. Specifically it compares analysis with direct access to the events (AOD and ESD data) to access mediated by different TAG-based event selections. We then compare different ways of splitting the processing to maximize performance.