Statistical Performance Analysis in a GPU

In this paper, we applied statistical analysis in order to evaluate the effect of parameters that affect the performance of a GPU-based parallel system. More specifically, we manually split the data to be processed into a number of trials that will perform inside the kernel while the rest of them are passed as parameters to calls of kernel. In addition, we used varying number of threads. For each combination that occurs by changing the above parameter values, we measured the speedup as the ratio of the CPU to the GPU code execution time. Also, we investigated GPU profiler's metrics to find out if there is any correlation with speedup. The performance evaluation was based on statistical analysis. Monte Carlo algorithms were used as benchmark, due to the high degree of parallelism they can incorporate.