Smart Server Selection: Enhancing QoE through a Budget-Aware Bandit in Meta Computing

Meta Computing aims to integrate cloud, edge, and terminal resources into a unified service environment. To make this paradigm practical in localized deployments, we develop a localized Meta Computing System (MCS) consisting of multiple Meta Computers (MCs), each aggregating nearby resources into a unified resource domain. A key problem is to select an appropriate MC for groups of routine tasks under uncertain, partially observed MC performance and a limited operating budget. Existing QoE-oriented resource-allocation methods do not jointly account for uncertainty in MC performance, heterogeneous operating costs, and diminishing marginal utility within a long-term subscription budget. We formulate the resulting MC selection problem as a budget-constrained multi-armed bandit with heterogeneous MC costs and a concave Quality of Experience (QoE) utility function. We then develop BUCB, an Upper Confidence Bound (UCB)-based policy that ranks feasible MCs by optimistic utility-cost efficiency. Under the stationary setting, we prove that BUCB achieves an expected regret upper bound that grows logarithmically with the total budget. Experiments in heavy-tailed and multi-modal environments under both stationary and non-stationary conditions, together with real-world validation on the WaterlooSQoE-IV dataset, demonstrate the effectiveness and robustness of BUCB relative to representative bandit baselines, achieving a 30-40% reduction in regret across various scenarios.