Towards Utilizing Remote GPUs for CUDA Program Execution