Nonlinear Network Programming on Vector Supercomputers: A Study on the CRAY X-MP

Abstract

The parallelism built into vector supercomputers raises several challenging issues for designers of optimization algorithms. We survey recent trends in parallel computer systems and study the impact of vector computing on nonlinear network programming. We propose a general framework for migrating fortran optimization software to a vector computer, and apply it in the context of two nonlinear network codes: NLPNETG, based on the primal truncated Newton algorithm, and GNSD, based on the simplicial decomposition method. We include computational experiments on a CRAY X-MP/24 system that tested the nonlinear network codes and compared the results with those of MINOS, a general purpose optimizer. Our experience indicates that vectorized codes can achieve significant improvements in performance (as much as 80% for primal truncated Newton), but achieve only modest improvements (15% for simplicial decomposition) for other algorithms.