Using Stream Processing to Improve the Speed of Hydraulic Network Solvers

Stream processing is a relatively new form of programming that employs a GPU (graphical processor unit) instead of solely relying on the CPU (central processing unit). Conventional programming uses the CPU to solve problems in a sequential manner. This is not very effective in solving simultaneous equations quickly. Recently, techniques have been developed that allows a GPU to be used to solve non-graphical mathematical problems. Standard GPUs being used today have over 100 threads compared to the 2 threads typically available to CPUs. Thus certain problems can potentially be solved at substantially higher speeds. The set of equations describing water network hydraulics often contains thousands of equations, which are nonlinear and have to be solved using an iterative process. This problem is potentially well suited to stream processing. However, the Cholesky method (a variant of Gaussian elimination tailored for symmetric positive definite matrixes) currently employed in network solvers is not well suited to stream processing, and an alternative matrix solver is required. The Jacobi iteration method has been identified as a suitable replacement, since each of its coefficients is treated independently of the other coefficients when solving the matrix. While work is still in progress, the potential for significant increases in the speed of solving water distribution hydraulic equations is already apparent.