Using multiple graphics accelerators to solve the two-dimensional inverse heat conduction problem

In this paper, we present a novel parallel algorithm implemented on graphics accelerators to solve the two-dimensional inverse heat conduction problem (estimation of the temporo-spatial heat transfer coefficients without any prior knowledge of the thermal boundary conditions) based on Particle Swarm Optimisation method. In the absence of analytical solutions, there are several heuristic methods to solve the problem, but the unacceptable high runtime (several days) makes these unsuitable for practical use. This paper presents the methods on how to adapt the original sequential algorithm to an efficient data-parallel one, keeping in mind the main features of graphics processing units (launching multiple threads on all multiprocessors, storing data in fast on-chip memory, eliminating warp divergence and memory transfer latency, using the host and device together, etc.). The achieved ∼45× speed-up (without any accuracy degradation) makes the heuristic methods suitable for practical use. Some of the proposed ideas are generally useable; therefore, this paper can be considered a step-by-step guide for researchers of other fields to speed-up general purpose calculations and evaluate the results.