SBIR Phase I:GPU Software for Sparse Linear Programming

This Small Business Innovation Research (SBIR) Phase I project will investigate the feasibility of developing a practical parallel implementation of the simplex algorithm for linear programming. As the demand to solve larger and more complex decision-making problems in all sectors of the economy increases, so does the demand for a parallel simplex algorithm for on-line and time-critical operations. Yet, parallel versions of this algorithm are currently unable to outperform appropriate serial codes for industrially-relevant problems that involve sparse large-scale matrices. The proposed research will explore a finer-grain parallelization than previously attempted for the simplex algorithm. As a result, the proposed implementation of the simplex algorithm will harness recent advances in graphics processing units (GPUs), while being capable of benefiting from any future serial algorithmic advances in the context of this algorithm. In addition to callable libraries and stand-alone executables, the implementation will be offered through cloud-enabled commercial services. The parallelization of the simplex algorithm has remained a major challenge in scientific computing. The proposed research will involve a systematic investigation of different combinations of factorization and pricing algorithms with regard to their performance on modern GPUs. If successful, the proposed work will result in an enabling technology for decision making in diverse areas across science and engineering and commerce. Linear programming is used extensively in applications ranging from data mining to airline scheduling, cancer therapy, engineering design, financial decision making, and logistics. As a result, the proposed work presents an unique opportunity to improve the competitiveness of the nation's commercial infrastructure.