Authors: Wenhao Gu, Xuchao Xie, Yujuan Tan, and Dezun Dongguan (National University of Defense Technology (NUDT))
Abstract: To reduce the overhead associated with transferring large volumes of data from storage to the CPU in data-intensive applications, data centers integrate data processing units (DPUs) to manage NVMe over Fabrics (NVMe-oF) remote storage access. However, current heterogeneous DPU and CPU system architectures for NVMe-oF face challenges of limited flexibility and suboptimal utilization of bandwidth and computational resources. In this paper, we present DNoF, a collaborative I/O processing stack for DPU and CPU in NVMe-oF systems, and implement it in a real-world environment. DNoF explores collaboration strategies between the DPU and CPU, achieving near-optimal system performance for NVMe-oF remote access under various workloads and delivering more than a 2x throughput improvement compared to DPU-only NVMe-oF.
Presentation: file