SHC-DP: Software–hardware collaborative in-network data placement for hybrid storage systems

Abstract

Hybrid storage systems increasingly adopt strategic data placement to deliver heterogeneous storage resources at low disk I/O latency for diverse applications. Modern adaptive algorithms with exceptional accuracy, adaptability, and scalability have gradually emerged as one of the mainstream algorithms for data placement. However, existing research predominantly focuses on exploring adaptive data placement algorithms with high placement accuracy, often neglecting the overhead introduced by these algorithms. It significantly compromises the end-to-end latency, a critical factor in determining the quality of service (QoS) for hybrid storage systems. To address this challenge, we introduce SHC-DP, a novel software–hardware collaborative in-network data placement framework. It is proposed to optimize the end-to-end latency for hybrid storage systems by leveraging modern SmartNIC technology. We first analyze the overheads associated with adaptive data placement in detail. Then, a comprehensive hardware/software co-optimization solution is proposed based on the observed insights. It incorporates algorithm acceleration, data transmission and processing optimization, computing and communication task overlapping, and concurrency optimization. Finally, we introduce detailed design and implementation for software/hardware modules and analyze SHC-DP’s performance bounds. Experimental results demonstrate that SHC-DP significantly optimizes end-to-end latency for hybrid storage systems, outperforming the state-of-the-art (SOTA) solution Sibyl by 1.54$\times$ $\sim 15.18\times$.