A Remote Attestation Protocol Oriented to the SuperPOD Structure

With the rapid development of artificial intelligence and machine learning in recent years, their complexity has also increased rapidly, and therefore the demand for high-performance computing has also increased rapidly. The traditional way of distributing training and inference work to a large number of servers will further increase the complexity of these tasks and slow down the process, which is particularly detrimental to those tasks that are time-sensitive. To solve these problems, many designs have emerged that integrate a large number of GPUs together, such as the DGX SuperPOD architecture launched by NVIDIA. However, as the number of GPUs in a server increases, security issues also arise. It is obviously too cumbersome to directly verify each GPU one by one. Therefore, we can refer to the swarm attestation mechanism of IoT devices and design a new remote attestation protocol based on structures such as SuperPOD that integrate a large number of GPUs together, so that users can quickly verify the security of a large number of GPUs in the SuperPOD structure, thereby improving the efficiency of remote attestation. This paper introduces the swarm attestation mechanism of the Internet of Things into the remote attestation of a multi-GPU cluster with a SuperPOD structure, proposes a remote attestation protocol based on the swarm attestation mechanism, and verifies in simulation that it is more efficient than simply expanding the ordinary one-to-one GPU remote attestation mechanism. At the same time, its analysis shows that the scheme can resist attacks such as forgery and replay.