 |
|
|
|
|
|
|
|
|
|
|
2004 SCIENCE NEWS
03.10.04 - Parallelization Tools Produce Dramatic Speedup
The performance of NIRVANA, a magnetohydrodynamic code used to study aspects of planet formation and accretion of gas onto the central star, was improved dramatically using the NASA Ames-developed CAPO (Computer-Aided Parallelization and Optimizer) tool. Within a week, the interdisciplinary team achieved a speedup of 3.8 on 4 processors and a speedup of 12 on 16 processors. For larger problem sizes the CAPO-generated code still sees further increase in parallel efficiency showing a speedup of 35 on 64 processors. The code was initially parallelized using the automatic parallelization feature of the SGI f90 compiler, but the generated code did not show significant speedup for more than four processors.
In order to investigate planet formation and to explain fundamental differences between extra-solar planets and our own solar system, scientist Adriane Steinacker studies the interaction of planets with the accretion disks from which they form. The fundamental numerical procedure employed by her NIRVANA code to solve the magnetohydrodynamical equations is the method of characteristics constrained transport. This computationally challenging task requires a large number of grid cells (eight million for a typical model), and many such models need to be run over a long time in order to cover a significant dynamical range.
The first step towards improving the efficiency was to obtain a performance profile for a couple sets of input data. The Paraver performance analysis system (from the European Center for Parallelism) was used for this purpose. The Paraver automated performance analysis component and the graphical user interface for performance trace examination helped to rapidly identify time-consuming routines with low parallel efficiency and to determine that there were large sections of the code which had not been parallelized or were parallelized inefficiently by the compiler. The source code was then fed to the CAPO computer aided parallelization tool. CAPO uses the extensive dependence analysis module of the ParaWise system, and, based on the information resulting from the analysis, inserts OpenMP directives into the source code. CAPO provides an extensive set of browsers to allow user interaction for improvements of the generated code. Dependence analysis and directive generation for NIRVANA took less than a minute. No user interaction was necessary. The CAPO generated code performed much better than the compiler parallelized code, due to fact that more of the time consuming loops were parallelized and the generated directives were placed more efficiently to avoid parallelization overhead.
Owner of the NIRVANA code, Adriane Steinacker of NASA Ames and UC Santa Cruz was very pleased with the results she saw: "I am very impressed by this stunning result that your tool brought on my code. This has the potential to help me catch up with my competitors, who came out with contradicting results, and who don't have the resources to look at this problem the way I can."
|
|
|
|
ARCHIVE
|
|
|
