That's why the sorting implementation is important. Unlike most CUDA programs which heavily relied on locality of data, sorting problem relies on the support of

That's why the sorting implementation is important. Unlike most CUDA programs which heavily relied on locality of data, sorting problem relies on the support of the global knowledge about key distribution. Even division-and-conquer methods such as qsort require a full scan over the dataset, which on CUDA architecture is painful. Duane's research concentrates on how to apply GPGPU architecture to a set of more general problems which traditionally relies on global access of data (thus, shared memory system). Disclaimer: I didn't follow his work closely, so my interpretation may be misleading.