Workload-Aware Replica Placement for Latency-Stable Cloud Services under Dynamic Traffic

Abstract

Replica placement has a direct impact on the latency and resource efficiency of distributed cloud services, particularly when request patterns vary across regions and time. This paper presents a workload-aware placement method that dynamically adjusts service replicas according to traffic intensity, inter-node communication cost, and observed tail latency. Rather than relying on fixed utilization thresholds, the proposed method estimates short-term workload changes from recent request traces and jointly determines replica count and placement location. A lightweight migration constraint is introduced to prevent excessive relocation during transient traffic fluctuations. Experiments are conducted using multi-region service traces under bursty, periodic, and gradually shifting workloads. The results show that the proposed method reduces P95 request latency and cross-node communication overhead while requiring fewer replica migrations than conventional threshold-based and load-balancing strategies. Additional experiments examine the effects of prediction errors and sudden workload changes, demonstrating that the method remains stable under imperfect workload estimates.