Charles Brooks and Pittsburgh Supercomputing Center scientist Bill Young have collaborated to exploit the parallel-processing power of the T3D for simulating the structure of proteins. The code they created is a specialized version of CHARMM (Chemistry at HARvard Macromolecular Mechanics), a major contributor to protein research and drug design since its development more than 15 years ago, by Martin Karplus, Brooks and others.
Several years ago, Brooks and Young created code that divides MD computations between a massively parallel system -- the Connection Machine, CM-2 or CM-5 -- and a traditional vector supercomputing system -- the CRAY Y-MP or C90. The heterogeneous (two different computing systems) version of CHARMM they developed tackles the intense computational demands of simulating a protein surrounded by water molecules. Their approach recognizes that a large part of these calculations involves interactions between water molecules, and these water-water interactions are inherently parallel -- computations for each water molecule are independent of the others. This part of the job can be carried out very efficiently on a massively parallel system at the same time as other interactions -- the protein with itself and protein-water -- are computed on the vector machine.
This distributed version of CHARMM ran successfully prior to availability of the CRAY T3D, establishing the potential of the approach, yet communications between different computing systems and other technical problems inhibited performance. With the T3D and C90 -- a closely coupled, single-vendor heterogeneous environment -- Brooks and Young resolved many of the technical problems. In tests on a wide range of structures, hetero-CHARMM on the T3D/C90 coupled system using 32 T3D processors completes its work two to three times faster than on the C90 alone.
"Molecular dynamics," says Young, "is an ideal application for parallel computers, and the T3D is more flexible and the processors faster than other systems we've worked on. In particular, there is a significant performance increase when scaling to a large number of processors. With algorithms we've developed to dynamically balance work across all processors, we're able to more efficiently use parallel computers and have much greater parallel speedups."
Young has begun using hetero CHARMM on the T3D/C90 to advance his work on alpha-helices, a fundamental motif of protein structure. "I'm pleased," says Young, "that we're past the technical problems now and running production code, actually doing science." His calculations investigate the thermodynamic changes that occur as the end of a helix begins to twist into another turn.
go back to the main screen