Merge Sort is a perfect choice for parallel sorting – it doesn’t require any synchronization. While counting sort requires some memory atomic operations – CUDA 1.0 doesn’t support Global Memory atomic operations – CUDA 1.1 or above supports ATOM while all nVidia Graphics Cards in our lab are CUDA 1.0… Also, I have to implement Merge Sort in the iterative way since CUDA doesn't support function recursion. What’s more, Merge Sort is an in-place sort which can help save a lot unnecessary memory space on GPU.
I would say it works well – but the current CUDA driver will kill the driver instance if the kernel function has been executing too long – only a few seconds. So it is the OS (windows) mechanism (watch dog) that forms the main obstacle. My professor said the latest CUDA driver will probably solve it.
Anyway, the performance comparison results are quite interesting. For small amount of testing number sizes, like 10k, 20k, CPU is much faster than GPU. But for 200k, 300k or even 1M, GPU will be much faster than CPU. – Parallelism is not silver bullet, isn’t it…