V.Palichik, E.Tikhonenko Joint Institute for Nuclear Research, Joliot Curie, 6, Dubna, Moscow region, 141980, Russia
email: eat@cv.jinr.dubna.su phone: 007 09621 67-553, fax: 007 09621 65-145
Abstract
With obtaining the SPP-2000 at the Joint Institute for Nuclear Research a few months ago, it provides the JINR specialists to use this supercomputer for parallel processing of jobs which require large computer and disk resources.
The JINR specialists take part in the experiments at the CERN, in particular in designing of the Compact Muon Solenid (CMS) setup which will be installed at the Large Hadron Collider (LHC). In the CMS experiment it requires detecting systems to have a very high spatial resolution (tens of microns) under high rates ( tens of nanoseconds). As result, data processing becomes significantly more complicated.
The search of an optimal transmission function for to measure a charge particle coordinate more precisely is one of important tasks for data processing for muon detectors. ( Now, at the construction phase of the CMS, we deal with data from Dubna muon detector prototypes; the data were obtained at the CERN SPS.)
This requires a minimization of some functional. It is a multi-parameter task (a total number of parameters is up to several thousands). The code includes operations on data (input, analysis and sorting of data) and a computational part (in particular, a lot of nested loops with calculations of sums for to minimize a functional). Such an ununiformity of the program does not allow to obtain a speed-up if we compile it with compiler options of optimization and parallelization only. The optimization has required to modify the algorithm and analyze a program structure for to define some parts of the program which are suitable for parallelization. So way, we have tested the variants with a usage of different directives (loop, task and region parallelization). The comparative results are presented.
The obtained results will be useful for further data processing and simulation in the CMS experiment and for the JINR specialists adopting their programs at SPP-2000 for parallelization.