VEUGM: Verifiable, Efficient and Universal GPU Performance Measurement in the Cloud

Due to the low-cost nature of cloud computing, many users deploy services on cloud platforms. However, malicious cloud platforms may offer hardware with lower performance than claimed, harming users' interests. Although users can run performance measurements on the cloud to verify the provided performance, existing methods are mostly designed for trusted environments and neglect the integrity of the measurement process. This oversight allows malicious cloud platforms to falsify measurement results easily. In this paper, we propose VEUGM, a verifiable, efficient measurement framework universal to various GPUs. This framework uses the TEE on the CPU side to generate measurement tasks and reliably verify the performance of cloud-based GPUs. Specifically, we devise an adaptive comprehensive performance measurement task that can accurately assess the performance of various GPUs with minimal overhead. Given the typically limited performance of TEEs, we further design a low-overhead measurement integrity verification algorithm to efficiently verify the correctness of measurement task execution within the TEE. Furthermore, we conduct theoretical analysis and extensive evaluations. The results on various GPUs demonstrate that VEUGM not only ensures the verifiability of the measurement task execution process but also reduces the required measurement time by an average of 82.63%.