SolarNowcast aims to forecast the solar irradiation from fisheye lens from webcam images. The sofware includes two components: “MotionEstimation” for the estimation of the dynamics from image data, and “Forecast” for the irradiation forecasting at short temporal horizon. A full performance evaluation on the Forecast code revealed a large optimization potential, both on serial (algorithmic, vectorization, memory usage) and parallel (openmp efficiency) levels. Follow-up optimization efforts improved the execution time, on the targeted production benchmark, by more than 2 times on the serial run, 4 times on 8 threads, and more than 5 times on 24 threads, and raised the scalability efficiency to 70% on 16 threads and 66% on 24 threads. A knowledge transfer then allowed the code holders to distribute and improve MotionEstimation, resulting in a factor 3 on 8 threads. This was obtained by improving the calculation part tenfold (x10), while an external minimizer part stays monothread. Further efforts will concentrate on this last part. The initial objectives of this work were to perform calculations in “real time” with respect to a given model and image acquisition rate. This milestone has already been accomplished on all tested machines, over 4 threads.
