Towards Efficient GPU Sharing on Multicore Processors