Research Assistants: K. Y. Wang, I.M. Martin, M. Sirbu
Post Doctoral Fellow: Z. Zhang
The effort to develop parallel algorithms for structural biology is the main focus of our research. The critical issues for high efficiency of the suite of molecular replacement programs [2] are data management [8] and load balancing [8]. Typical problems involve 3-D meshes with up to 400 million grid points. The computation for grid points located within the molecular envelope requires electron density averaging of all the points related by noncrystallographic symmetry. The points related to a grid point by a group of symmetry operators are scattered throughout the entire mesh; therefore, data management is critical. Data is either cached in the local memory of all the compute nodes or data servers, or dedicated compute nodes with a substantial amount of memory are assigned to the task of caching the data. In the first case, data delivery is based on interrupt driven messages; in the second case, polling schemes and asynchronous communication are used.
The computation is nonhomogenous because the subset of grid points within the molecular envelope requires intensive computations while those located in the solvent or the nucleic acid require very light computations. The programs partition the data into blocks called data allocation units (DAUs), which are tagged by the amount of work required to process the data in the DAU. The programs use a sequencing algorithm to access the DAU, which adapts itself to ensure near-minimum communication cost. The transformation required between phases of the computation is now optimized and uses a global, in-place exchange algorithm to minimize contention on the communication links.
Another facet of our research consists in developing intelligent user interfaces and problem-solving environments [4,5,7] to facilitate access to supercomputers for scientist working in different areas. The structural biology problem-solving environment allows transparent access to supercomputers with different architectures.
Descriptions of three projects related to this research, Bond, a parallel virtual environment, Tonitza a visualization package for Structural Biology Data and MR, Molecular Replacement Programs is available on the World Wide Web (http: //www.cs.purdue.edu/people/dcm).
Acknowledgments
This work is supported in part by NSF grants CCR-919388, BIR-9301210, and MCR 9527131.