Hierarchical Self-Optimizing AI Models For Efficient Resource Allocation In Cloud Computing Management Applications

Abstract

Cloud computing infrastructure needs to dynamically distribute compute, memory, and networking resources among thousands of tasks and servers in varying loads, but current autoscaling and placement systems are typically based on pre-defined threshold policies or just one monolithic reinforcement learning agent which has to make decisions about global capacity budgeting, cluster-level allocation, and server-level task scheduling all at once in the same unifying policy, a method that is not scalable in the face of increasingly large state-action spaces. In this paper, we introduce a Hierarchical Self-Optimizing Resource Allocator (HSORA), a hierarchical approach to cloud resource management with three layers, a global layer allocating capacity budgets between clusters by proximal policy optimization, a mid-layer for per-cluster VM/container placement and autoscaling using actor-critic algorithms, and a local layer managing task scheduling and power management on individual servers, coupled with a self-optimization monitoring system to detect when a change in workload patterns has caused SLA and cost efficiency violations and then retrain only the relevant layer. The performance of the framework has been tested on statistically modeled cloud workload data based on Google cluster trace and Azure Public Dataset in four different scenarios of steady state, bursty, multi-tenant contention, and workload pattern change. In comparison to the performance of a static threshold-based autoscaling method and a conventional single agent Deep Reinforcement Learning model, the proposed HSORA has managed to cut down the SLA violation ratio from 18.73 and 12.92 percent to 5.40 percent while improving resource utilization efficiency to 92.4 percent.