Optimizing Secure Data Transmission in GPU-Accelerated Collaborative Computing via Metadata Reconstruction

As the use of GPUs in collaborative computing environments increases, the demand for memory capacity continues to grow. The introduction of Compute Express Link (CXL) technology enables GPUs to access extended memory but also introduces new security challenges. Expressly, data transmission between secure memory regions incurs overhead due to ciphertext conversion, driven by independent security meta data for each memory region, which results in additional memory access overhead and degrades performance. Existing solutions alleviate this issue by sharing security metadata, thus reducing overhead from additional memory accesses. However, transmitting security metadata still consumes significant bandwidth, limiting overall system performance. To address this limitation, we propose a data transmission optimization method based on secure metadata reconstruction tailored to the unique characteristics of GPU multi-memory architectures in collaborative computing. Our method minimizes the amount of security metadata transmitted and reconstructs complete metadata at the receiver's end. This approach significantly reduces bandwidth usage and enhances system performance. Experimental results demonstrate that our method outperforms the Salus approach, achieving an 11.0% improvement in IPC (Instructions Per Cycle) performance.