  |
|
|
|
As we discussed last week, understanding the mechanisms of
protein folding is an extremely difficult problem from a
computational point of view. This is due in large part to
the extremely intensive nature of such calculations. But
as computers get faster and algorithms are revised to take
better advantages of the resources available, it is becoming
feasible to study such systems. One group of scientists
working in the Department of Pharmaceutical Chemistry at
the University of California at San Francisco has made
significant advances in this particular area.
The first piece of the puzzle was the availability of a Cray T3D and a Cray T3E, both with 256 processors, at the Pittsburgh Supercomputer Center. Presented with these resources, the Kollman group first went to work optimizing the code to be used in the simulation, namely the molecular dynamics capability of the AMBER package. The first effort resulted in a 70% performance increase in single-processor mode of the software. (It's interesting to note that there is often performance enhancement that can be gained in a lot of scientific software, but it either remains uncovered, or is considered to be unworthy of the coding effort and time needed to achieve it. When the simulation time savings can be multiplied over many processors, however, the "insignificant" speedups often can make the difference between being able to study a system and not being able to!) Additional work was done to enhance the load balancing achieved when multiple processors are used, and the total 256-processor speed enhancement was a factor of six. With this souped-up application, the first 0.2 microseconds STILL required 40 days to complete ... using all 256 processors of the T3D! The remaining 0.8 microseconds of simulation were completed using 256 processors on the T3E, which is about four times faster than the T3D. This leg of the study lasted another two months. The study produced some interesting results and provides some insight into the mechanism of protein folding. To read more about this study, you can visit http://www.psc.edu/science/Kollman98/kol... The take-home lesson here is that there is a lot that can be done by using or enhancing existing methods on existing machines if the problem is approached in the right way, but there is always a need to push Go To Page: 1 2
The copyright of the article Parallel Protein Folding
in Scientific Computing is owned by Adam Hughes. Permission to republish Parallel Protein Folding
in print or online must be granted by the author in writing.
|
|
|
|
