Exploring Dynamic Load Balancing in Real Valued Indexed Arrays for Short Range Molecular Dynamics Calculations

This paper explores some of the issues in the improvement of the performance of Real−Valued Indexed (RVI) arrays using dynamic load balancing. RVI arrays are a possible extension to FORTRAN which allow the programmer to express spatial problems in an intuitive manner allowing the compiler and run−time environment flexibility in the actual storage layout of the data. While RVI arrays in general are expected to have broad application, this paper focuses on the use of RVI arrays to solve dynamic molecular dynamics problems. A prototype RVI array library has been implemented and tested on a number of workstations systems, scalable shared memory BBN parallel processing systems, and the SGI parallel processing systems. The dynamic load balancing techniques used in RVI are a combination of balancing computations, data locality, and tasks. This paper is best read on−line using a World−Wide Web browser such as NCSA Mosaic browser. When this paper is viewed on−line, all of the figures are in color and the animations can be viewed. The URL for this paper is: http://clunix.msu.edu/~crs/papers/rvi_94/