A Dynamic Protection Mechanism for GPU Memory Overflow

Part 1: AcceleratorInternational audienceGraphics Processing Units (GPU) are widely used to accelerate computation in many applications such as autonomous vehicles, artificial intelligence and healthcare. However, most existing researches just focus on the performance but ignore the security issues of GPUs. In this paper, we design an efficient mechanism to dynamically monitor GPU heap buffer overflow by using the CPU. Concretely, we first analyze the specific requirements of GPU memory allocation. Second, in order to realize the monitoring from the CPU, we map the allocated device memory to the host-side. Third, the dynamic monitoring of buffer overflow is implemented based on the mapped memory. Our results show that it is feasible to protect the GPU memory from the CPU side. Our work can improve the efficiency of GPU memory allocation and increase the security at the same time. By offloading the detection of buffer overflow to the CPU, the performance of GPU kernels will not be affected significantly