GPUTeraSort: High Performance Graphics Coprocessor Sorting for Large Database Management

Sorting large datasets is often limited by I/O bandwidth in terms of memory and disk. The traditional von Neumann architecture results in high cache misses within L1-L3 cache levels. The authors realized that the highly parallelized processors and the fast memory interconnects inside commodity GPUs can help to work around the limitations that arise when sorting is done solely on the CPU. Specifically, the limitation is that most sorting performance is limited by cache misses. Although main memory sizes have gotten larger, so that disk I/O is reduced, memory access itself is a bottleneck. Utilizing the GPU is advantageous in that it may have 10 times the bandwidth for memory that a CPU has.