Efficiently Using a CUDA-enabled GPU as Shared Resource