AI-Enhanced Stage-Aware Deadline Division for Secure Multi-Cloud Resource Management and Performance Prediction

Abstract

Using the identical traces and capacity controls described in Section 4, SBSAD decreased the end-to-end turnaround time and increases the satisfaction rate of meeting deadlines, while reducing the number of late or dropped tasks and preemptions. These gains stemmed from stage-based deadline bucketing and a single scalar priority that considered urgency, security, dependency, and predicted runtime. This was in contrast to DRL-RSM, where reward shaping and global queues make deadline pressure less explicit. Unlike STGC-SM that relied on static graph features or offline prediction, STGC-SM directly consumed online multitasking prediction (runtime/energy/fault risk) within the scheduler. Therefore, its response to short-term load changes was slower. Moreover, by preempting cost feedback for cross stage promotion, it avoided the convergence overhead observed in the sudden arrival of MARL-RMM. In practical cloud stacks, integration was straightforward: In Kubernetes, a scheduler framework plugin or extender could replace the native scoring with a scalar priority while applying GCN-derived feasibility masks in the filter phase. The predictor was deployed as a telemetry-driven sidecar. Equivalent behavior in OpenStack-like platforms followed from custom host filters and weights. Operational interpretation revealed that improved accuracy reduced queue build-up and deadline violations for batch and ML workflows in shared clusters. It also lowered interference in isolation-critical, multi-tenant SaaS and stabilized edge-to-cloud video and ETL pipelines during short-term surges. Limitations remain: Reliance on predictor quality and stable telemetry can bias prioritization. The CloudSim environment simplifies network and storage contention compared to production systems. Preemption cost and fairness are modeled beyond deadline satisfaction at a coarse level. These factors define low-risk deployment zones and indicate where additional engineering hardening or A/B trials are recommended.