Supercomputers derive their computational performance from faster processors as well as innovations in their architecture. To take advantage of the vector processing capabilities of supercomputers, such as the cray x-mp series, it is necessary to modify the code to enhance its vector processing performance. These modifications can range from simple localized recoding of existing mainframe codes to devising new algorithms with the hardware's architecture in mind. In this paper, codes for the solution of two network equilibrium assignment problem formulations (frank-wolfe algorithm for the single-class user equilibrium problem and the diagonalization algorithm for multiple user classes with asymmetric interactions) are vectorized and tested on a cray x-mp/24 supercomputer. Only local vectorization by limited recoding of existing programs is performed. Guidelines are given for this purpose, and their application to the assignment codes is illustrated. The computational tests performed indicate an improvement in execution time of about 70 to 80% of the modified code relative to its unvectorized performance on the cray supercomputer. Execution of the vectorized code on the cray is about 22 times faster than the execution of the unmodified code on a mainframe computer. The significance of the results for research and practice is also discussed. This paper appears in transportation research record no. 1251, Transport supply analysis.
Abstract