A systematic evaluation of accelerating indoor airflow simulations using cross-platform parallel computing

With new advances in computer hardware and software, users now have widespread accessibility to multicore devices inside personal computers making it feasible for fast indoor airflow simulations. Some exciting preliminary results of a cross-platform parallel computing framework OpenCL using specific hardware were reported. However, those results are largely based on two hypotheses: 1. OpenCL code on all devices will generate the same results; 2. On the same device, running in parallel with multiple processors will be faster than running in sequential with a single processor. This study attempted to evaluate these two hypotheses by systematically studying the accuracy and computing speed of OpenCL for indoor airflow simulations. A Fast Fluid Dynamics (FFD) code was selected as an exemplar indoor airflow simulation program. To compare the cross-platform ability of OpenCL, the evaluation was performed using four Graphics Processing Units (GPUs) and five Central Processing Units (CPUs) from three manufacturers, with different degrees of computing capability and mounted on two operating systems. The test subjects were evaluated using four case studies consisting of various indoor airflows. A sequential FFD code programmed in C and a Computational Fluid Dynamics (CFD) program were first used to perform the case studies and generate numerical benchmarks. The comparison of the numerical simulation results with experimental data showed that CFD and FFD can predict the studied flows with averaged relative errors of 9.99% and 11.30%, respectively. Afterwards, the accuracy and speedup of the OpenCL code was compared with numerical benchmarks. Although the OpenCL code on the CPUs generated identical numerical results, the OpenCL results from the GPUs were slightly dissimilar. This is likely due to varying interpretations, by the manufactures, of an Institute of Electrical and Electronics Engineers standard. Depending on the hardware, the speedups of the OpenCL code varied from 0.7 to 4.2 times on the CPUs and 5.1-129.3 times on the GPUs. The slowdown of computing speed happened when running OpenCL on a two-core CPU in a Windows Operating System using the Boot Camp on a Mac computer. Finally, a separate study on the relationship between speedup and global work size showed that a speedup of 1139 can be achieved when using an AMD FirePro W8100 GPU.