Optimized GPU Sorting Algorithms on Special Input Distributions

We present a high performance graphics processing unit (GPU) sorting algorithm ISSD (Improved Sorting considering Special Distributions) implemented with the Compute Unified Device Architecture (CUDA). The ISSD focuses on two aspects to improve parallel sorting performance. One is how to decompose the sorting tasks into independent and balanced subtasks which can then be easily distributed to thousands of threads to realize the concept of “parallel sorting” as well as to efficiently explore the power of GPU. The other one is how to take advantage of special data distributions to further optimize the algorithms and improve their performance. The algorithm is redesigned based on our previous general data distribution version and optimized both on general implementation methods and special input distributions. Experimental results show that for the general data distribution inputs, the ISSD outperforms the existing parallel sorting algorithms by about 10% in performance due to its practical optimization in implementation; and for three special data distribution inputs, the ISSD outperforms the existing algorithms by more than 40% due to its special optimization based on the data distributions. Therefore, the algorithm is viable and efficient when dealing with specific data distributions.