High-Performance Computation in Biomolecular Modeling

Thursday, January 23, 1997 - 9:30am - 10:30am
L. Ridgway Scott (University of Houston)
High-performance computation offers both challenges and opportunities for biomolecular modeling. Numerous standard codes are now available on a wide range of platforms. These codes are being used to do simulations of biologically important systems that are an order of magnitude larger than previously possible, and for a significantly smaller cost. We highlight some of the efforts carried out in our group as part of the National High-Performance Computing and Communication Initiative. Not only is this work having a significant impact on the development of biomolecular science, it is also having a substantial influence on the development of future high-performance computing platforms.

We describe the development of some parallel iterative techniques for solving boundary value problems for elliptic partial differential equations. Using domain decomposition techniques, we have modified standard sequential iterative techniques to obtain effective parallel methods with minimal code restructuring. We contrast implementations on distributed-memory and shared-memory scalable parallel processors. We describe the use of two different programming paradigms, one involving explicit parallelism in a distributed-memory model and the other utilizing simple loop decompositions in a shared-memory model. Our primary conclusion is that parallel computing on existing commercial parallel supercomputers makes it routine to do three-dimensional modeling of biomolecular systems.

We also describe similar successes in parallelizing and using existing codes for molecular dynamics. One of these involves a minimal change to the original code yet provides substantial parallel performance up to nearly a hundred processors. Another more ambitious project has developed a more scalable version which demonstrates acceptable performance on several hundred processors. This code is being used to study a full dimer of acetyl-cholinesterase in solution, involving over 130,000 atoms in the simulation.

Finally, we will mention recent work parallelizing molecular imaging codes. This is allowing electron microscope data to be reconstructed to a much greater accuracy than possible before. The computational algorithms that are widely used have novel data access patterns that pose interesting challenges for distributed-shared-memory systems. These emerging parallel supercomputer systems are expected to dominate the market in the future, and imaging algorithms provide an important new source of guiding experience that can help in making critical design decisions for novel computer architectures.