The rapid growth of data-intensive applications has driven the widespread adoption of heterogeneous object storage (HOS) systems, which tier data across high-performance and high-capacity storage devices. However, the widening performance gap across tiers exacerbates service-level objective (SLO) violations. Existing prediction and scheduling approaches often overlook device heterogeneity and cross-device spatial dependencies (e.g., performance coupling and interference), which can turn localized congestion into systemic bottlenecks. To address these challenges, we propose STGraph3PO, a spatio-temporal graph learning framework with three key innovations. First, we model HOS systems as spatio-temporal graphs and employ a graph attention network to capture device heterogeneity by adaptively weighting node interactions. The learned spatial representations are then combined with gated recurrent units to predict per-device queueing delays. Second, we introduce virtual edges to enhance attribute propagation among correlated nodes, thereby accelerating model convergence and enabling efficient online adaptation. Third, we design an SLO-aware control strategy that proactively identifies high-risk requests and mitigates violations via device-aware prioritization and intra-queue scheduling. Experimental results show that STGraph3PO reduces the SLO violation ratio by up to 1.49 percentage points in 5 out of 6 scenarios and improves latency by 12.7-26.3% from average to P999 levels.
Zhang et al. (Fri,) studied this question.