Adaptive Cloud Resource Allocation using Deep Reinforcement Learning

Abstract

Allocating cloud resources — deciding which physical machine should host each arriving task, virtual machine, or container — is a combinatorial, online, multi-resource problem that classical bin-packing heuristics (first-fit, best-fit) solve only crudely: they ignore workload structure, over- or under-consolidate, and leave hosts fragmented. This paper formulates adaptive resource allocation as a Markov decision process and learns the allocation policy with deep reinforcement learning (DRL). The scheduler observes the heterogeneous CPU/memory/GPU utilization of a machine pool and the resource profile of the next task, and chooses a placement that maximizes a composite reward trading utilization against energy, SLA violations, and resource fragmentation. We give the MDP formulation, the multi-resource utilization and reward definitions, and the value objective; survey the algorithm landscape (DQN and its Rainbow extensions, A3C, and multi-agent variants); and discuss GPU fragmentation and consolidation as special cases. Illustrative results, consistent with the published literature, show that a learned policy consolidates onto fewer active hosts, raising utilization and cutting energy while reducing SLA violations relative to heuristic baselines. This paper is distinct from horizontal auto-scaling: it concerns where to place work on a fixed pool, not how many replicas to run. Public sources are cited throughout.