Speaker
Dr
Isidro Gonzalez Caballero
(Instituto de Fisica de Cantabria (CSIC-UC))
Description
A typical HEP analysis in the LHC experiments involves the processing of data
corresponding to several million events, terabytes of information, to be analysed in
the last phases. Currently, processing one million events in a single modern
workstation takes several hours, thus slowing the analysis cycle. The desirable
computing model for a physicist would be closer to a High Performance Computing
one where a large number of CPUs are required for short periods (of the order of
several minutes). Where CPU farms are available, parallel computing is an obvious
solution to this problem. Here we present the tests along this line using a tool for
parallel physics analysis in CMS based on the PROOF libraries. Special attention has
been paid in the development of this tool to modularity and easiness of usage to
enable the possibility of sharing algorithms and simplifying software extensibility
while hiding the details of the parallelisation.
The first tests performed using a medium size (90 nodes) cluster of dual processor
machines on a typical CMS analysis dataset (corresponding to root files for one
million top qurk pairs producing fully leptonic final state events distributed
uniformly among the computers) show quite promising results on scalability.
Primary author
Dr
Isidro Gonzalez Caballero
(Instituto de Fisica de Cantabria (CSIC-UC))
Co-authors
Mr
Daniel Cano
(Instituto de Fisica de Cantabria (CSIC-UC))
Dr
Javier Cuevas
(Departamento de Física, Universidad de Oviedo)
Mr
Rafael Marco
(Instituto de Fisica de Cantabria (CSIC-UC))
