Leveraging Quantum Deficit Modeling for Scalable Multi-Tenant Resource Allocation on Advanced Cloud Datacenter Platforms

Multi-tenancy is fundamental to cloud storage, but efficient resource sharing with predictable performance remains challenging because SSD internals are opaque and shared-resource contention increases tail latency. As a result, to meet latency targets, existing host-side approaches rely on conservative over-provisioning, reducing utilization. In this paper, we present DILATON, a closed-loop Quality of Service (QoS)-aware scheduling framework for multi-tenant NVMe (Non-Volatile Memory Express) SSDs (Solid State Drives). DILATON is motivated by the observation that QoS degradation arises from both back-end contention and the mismatch between issued and completed requests across domains. We leverage the submission–completion gap and resulting per-domain percentile-latency deficits as implicit feedback signals for dynamic request regulation. We develop a contention-aware QoS model that captures die-level collision probability and relates queue depth to bandwidth and latency. Using this model, DILATON adjusts per-domain queue depths to meet QoS targets, transforming conventional open-loop NVMe scheduling into a closed-loop framework that aligns host-side demand with SSD’s back-end capacity. We implement a multi-dimensional quantum-deficit scheduler with credit-based admission control and a feedback-driven controller for dynamic adaptation. We use a real-system testbed to evaluate open-loop control and demonstrate the submission–completion gap that motivates feedback, while closed-loop policies are evaluated in simulation. QD Control uses the submission–completion gap to balance resource allocations across domains, while QoS Control uses percentile-latency deficits to satisfy latency targets. The results show that DILATON meets its tail-latency targets across all priority classes in the configurations evaluated, allocating capacity adaptively rather than through static over-provisioning.