Experience from a pilot based system for ATLAS

Abstract

The PanDA software provides a highly performing distributed production and distributed analysis system. It is the first system in the ATLAS experiment [1] to use a pilot based late job delivery technique. This paper describes the architecture of the pilot system used in PanDA. Unique features have been implemented for high reliability automation in a distributed environment. Performance of PanDA is analyzed from one and a half years of experience of performing distributed computing on the Open Science Grid (OSG) infrastructure. Experience with pilot delivery mechanism using Condor-G [2], and a glide-in factory developed under OSG will be described.