Abstract
<title>Abstract</title> <p>Hyperconverged infrastructure (HCI) colocates compute, storage and networking on the same physical nodes, but the defaultKubernetes scheduler operates reactively and is un-aware of the underlying Ceph storage state, which causes resourcecontention, load imbalance and elevated I/O latency under mixed workloads. Existing studies have addressed scheduling andstorage optimization in isolation; an integrated framework that couples short-horizon workload prediction with storage-awareplacement on a single HCI control plane has not been reported. We propose a predictive and adaptive scheduling model thatfuses an exponentially weighted moving average (EWMA) demand forecaster, adaptive pod placement, dynamic resourcere-allocation, and Ceph-aware storage decisions into one closed-loop controller, operating entirely through the standardKubernetes and Pro-metheus APIs without modifying core components. The model was implemented and evaluated on areal three-node hyperconverged testbed running Proxmox VE 9.1.4, Ceph 19.2.3 and Kubernetes 1.29.15, under a mix of CPU-intensive (stress-ng), I/O-intensive (fio) and transactional (YCSB) workloads. Compared with the default Kubernetesscheduler, the proposed model reduced CPU load imbalance by 79.4 percent (from 53.29 percent to 10.98 percent), loweredmaximum node utilization from 95.64 percent to 67.41 percent, raised minimum utilization from 39.73 percent to 56.33 percent,and reduced average Ceph OSD apply latency from 1.11 ms to 1.00 ms. The results show that cross-layer architectural integration yields greater performance gains than algorithmic complexity alone.</p>