ARDA: I/O Scheduler for Heterogeneous Workloads Co-located on Ultra-low-latency SSDs

Abstract

Ultra-low-latency (ULL) SSDs enable cloud service providers to co-locate latency-sensitive services and throughputoriented background jobs on the same machines. However, their microsecond-scale latency creates a scheduling dilemma: conventional I/O schedulers introduce visible overhead, while disabling scheduling removes the kernel's ability to enforce workload-specific Quality of Service (QoS). This paper presents ARDA, an Adaptive Request Dispatching Algorithm for QoSaware I/O scheduling on ULL SSDs. ARDA dynamically assigns dispatch budgets based on workload QoS requirements inspired by real-time CPU scheduling concepts. Unlike conventional schedulers that rely on request types or relative priorities, ARDA regulates the amount of I/O each workload may inject into the device, thereby satisfying latency targets while preserving throughput and avoiding hardware queue overload. Evaluation on an emulated ULL SSD shows that ARDA reduces average latency by $\mathbf{2 2. 9 6 - 4 2. 9 5 \%}$ and maximum latency by $\mathbf{3 6. 2 6 - 4 4. 8 7 \%}$ compared with Kyber under mixed workloads. Under writesaturated conditions, ARDA reduces worst-case latency by up to 59% compared with the none scheduler. These results show that target-aware scheduling can provide predictable QoS for colocated workloads on ULL SSDs without sacrificing throughput.