Identifying Performance Bottlenecks In A Radiative TransferApplication

The Hydrolight software is a well known radiative transfer solver, written in Fortran 77, intended for hydrologic optics problems. It was developed by C. D. Mobley and it is based on an invariant imbedding methodology that yields a system of non-linear ordinary differential equations, the Riccati equations, which are solved by routines that implement numerical methods. This work proposes an innovative strategy for identifying performance bottlenecks and optimising a code using both standard profiling tools and accessing specific cpu registers and a time counter. The execution profile in a RISC worst at ion shows that most of the processing time is spent in solving Riccati equations, what is accomplished by a Runge-Kutta algorithm. The time-consuming routines were identified and a timing routine based on system wall time was used to analyse which are the critical parts of these routines. The code was then ported to an Intel monoprocessor, running under Linux, in order to be monitored by cpu event counters. Based on this detailed data, classical optimisations were performed in the related routines and the code was re-run and its performance re-evaluated. The optimised code was ported back to the RISC workstation and benchmarked again. Another step was to port the time-intensive routines to Fortran 90 arid check performance enhancements. We are currently parallelising the new code on a Beowulf-class system based on Pentium-Pro processors. The main technique that we are employing in this task is the insertion in the Fortran 90 code of High Performance Fortran (HPF) directives yielding a code suitable for both shared memory or distributed memory parallel systems. Applications of High Performance Computing in Engineering VI, C.A. Brebbia, M. Ingber & H. Power (Editors) © 2000 WIT Press, www.witpress.com, ISBN 1-85312-810-4