隨著 Zoned Namespace（ZNS）SSD 的日益普及，記憶體管理面臨新的挑戰，特別是在處理必須遵循順序寫入限制的 swap 流量時。 在本研究中，我們提出了一種 群組感知（group-aware） 的 swap 分配機制，透過在動態形成的節點群組（每個群組對應至不同 die）之間，以循環（round-robin）方式發出固定大小的區塊寫入。此策略能同時發揮裝置內部平行性、避免寫入熱點，並確保遵守裝置限制。 我們在主機端實作了此機制，並使用具有不同圖結構的異質圖神經網路（Heterogeneous Graph Neural Network，HetGNN）訓練工作負載進行評估。實驗結果顯示，我們的設計在某些情況下可將 P99 尾延遲（tail latency）降低高達 44%，同時實現更均衡的 die 級別寫入分佈並減少資源爭用。與基準策略相比，本方法在複雜記憶體存取模式下提供更平穩且可預測的 swap 效能，突顯了其在現代 ZNS 儲存系統中管理多樣化與動態工作負載的有效性。 ;The increasing adoption of Zoned Namespace (ZNS) SSDs introduces new challenges for memory management, particularly when handling swap traffic that must adhere to sequential zone write constraints. In this work, we propose a group-aware swap allocation mechanism that issues fixed-size chunk writes in a round-robin fashion across dynamically formed node groups, each mapped to separate dies. This strategy exploits internal parallelism while avoiding write hot-spots and preserving device constraints. We implement this mechanism at the host level and evaluate it using Heterogeneous Graph Neural Network (HetGNN) training workloads with varying graph structures. Experimental results show that our design significantly reduces P99 tail latency—by up to 44\% in some cases—while achieving more balanced die-level write distribution and minimizing contention. Compared to baseline strategies, our method delivers smoother and more predictable swap performance under complex memory access patterns, highlighting its effectiveness in managing diverse and dynamic workloads within modern ZNS-based storage systems.
