PARALLEL PROCESSING OF LARGE DATA SETS IN PARTICLE PHYSICS

The analysis of the LHC data aims to minimize the vast amounts of data and the number of observables used. After slimming and skimming the data, the remaining terabytes of ROOT files hold a selection of the events and a fiat structure for the variables needed that can be more easily inspected and traversed in the final stages of the analysis. PROOF has an efficient mechanism to distribute the analysis load by taking advantage of all the cores in modern CPUs through PROOF-Lite, PROOF Cluster or PROOF on Demand tools. In this paper we compared performance of different methods of file access (NFS, XROOTD, RFIO). The tests were done on Bucharest ATLAS Analysis Facility.