A Framework of GPU Transparent Offloading for Cloud Robotics

The limited computational resources on the robot side make it impossible to deploy many robotics applications to real-world scenarios yet. Although cloud GPU servers can provide ultra-high computing power, these commercial applications are usually not open source, so there is a need to offload GPU computation to the cloud GPU for computation without understanding the application code and without intervening in the operation of the application, i.e., to achieve GPU computation transparent offloading. In addition, an automatically recorded replay mechanism is proposed for the framework to reduce the network transmission overhead during the process of offloading. Through numerous experiments, the NVIDIA GPU computation transparent offloading system brings positive combined gains of 246 ms and 778 ms for the multi-person pose recognition application KAPAO and the indoor stereo environment reconstruction application NICESLAM, respectively; for the sweeper robot path planning application CALVIN and the machine car and human visual interaction casual game Stylemesh brought negative combined gains of -21 ms and -717 ms, respectively. In the end, the experiments prove that the our proposed GPU computation transparent offloading framework is correct and the engineering is feasible. However, there are positive and negative gains from the data, indicating that some scenarios are applicable to computation offloading and some scenarios are not applicable to computation offloading.