Francisco Almeida, Evencio Mediavilla. Alex Oscoz
and Francisco de Sande
Dpto. de Estadistica, I.O. y Computacion
Universidad de La Laguna, La Laguna, Spain
email: sande@csi.ull.es
This work collects the experiences of a collaboration between researchers
coming from two different fields: astrophysics and parallel computing.
We deal with different approaches to the parallelization of
a scientific code that solves an important problem in astrophysics, the
detection of supermassive black holes.
To accomplish it, light curves of Quasi-stellar Object (QSO) images
must be analytically modeled. The scientific aim is to find the values that minimize the error between this theoretical model
and the observational data according to a chi-square criterion. The robustness of this procedure depends on the sampling size 
.
However, even for relatively small values of , the determination of the minimum takes a long
time of execution since 
starting points of the grid must be considered.
We will show that parallelization can reduce the total time
of execution, allowing to greatly increase the sampling and, consequently, to
improve the robustness of the evaluated minimum.
The sequential code is best suited for different parallel implementations varying from message-passing (MPI) to shared-memory (OpenMP) codes and hybrid solutions combining both (MPI+OpenMP). One of the aims of this work is to obtain the maximum reduction in the execution time for the original sequential code. With this goal in mind, we have also to consider that the amount of time invested in developments is a heavy limitation for the usual non-parallel expert scientific programmers. To balance both objectives we devise the necessary developments comparing pure MPI and OpenMP codes and mixed MPI+OpenMP programs also. The parallel approaches have been tested and a broad computational experience on a ccNUMA SGI Origin 3000 will be presented. The conclusions and the knowledge acquired with our experiences establish the basis for future developments on similar codes.