A Fairness-Aware Management for Multi-Tier Storage Systems

Modern cloud storage systems serve multiple tenants with diverse I/O behaviors and make fairness under shared resource contention a critical requirement. To mitigate the random-write limitations of high-capacity Shingled-Magnetic-Recording (SMR) disks, hybrid multi-tier architectures are widely adopted. However, ensuring fairness in such systems remains challenging because cache optimization and SMR cleaning mitigation constitute fundamentally different bottlenecks that evolve over distinct time scales. This paper presents a fairness-aware management for multi-tier storage systems by integrating a DRAM cache, a flash-based persistent cache, and an SMR backend. By explicitly decoupling short-term cache contention from long-term cleaning interference, the proposed method applies tier-specific priority management to regulate hit rate degradation and cleaning slowdown in a coordinated manner across the storage hierarchy. Experimental results on ten mixed workloads show that the proposed method improves integrated fairness by up to 4.59 × while reducing average response time and P99.9 long-tail latency by up to 11.4% and 15.2%, respectively.