PanDA's Role in ATLAS Computing Model Evolution

During Run 1 at the Large Hadron Collider from 2009-2013, the ATLAS experiment successfully met the computing challenge of accumulating, managing and analyzing a volume of data now exceeding 140 PB, processed at over 100 sites around the world, and accessed by thousands of physicists. This accomplishment required nimbleness and flexibility in the distributed computing infrastructure, both hardware and software, as the operational computing model evolved during the run based on experience. A critical enabler for this evolution was PanDA, the ATLAS workload management system used for production and distributed analysis. PanDA's capabilities were utilized and extended to dynamically and intelligently distribute data and processing workloads across ATLAS resources based on data popularity and resource availability, thereby 'flattening' an originally hierarchical computing model, in order to use resources more efficiently. A new round of PanDA development now taking place will continue to evolve the model for better utilization of distributed resources. Examples include dynamic partitioning of workloads based on available resources; flexible data locality requirements enabled by direct data access across the WAN; and a new Event Service approach to deliver fine grained processing. This presentation will describe the Run 1 experience and new developments under way for Run 2 at the LHC.