應用於行動裝置之混合式記憶體交換系統

Abstract

行動裝置在現今複雜的應用程式日益增長的記憶體需求與有限的實體DRAM之間，持續面臨著挑戰，這導致低記憶體終止機制 (LMK) 頻繁地終止程序 (process)，從而降低了多工處理的使用者體驗。儘管先前如 Marvin 和 Fleet 等物件級交換框架，已試圖透過優化回收速度或熱啟動延遲等單一效能指標來應對此問題，但它們在複雜且不可預測的工作負載下，仍可能表現出急遽的「效能懸崖」。 本論文提出 PhantomSwap，一個為行動裝置設計的混合式記憶體交換系統，其目標在於全面地改善系統流暢度與韌性。PhantomSwap 的核心創新是一個由智慧型「統一交換控制器」所管理的反應式三層式記憶體階層（DRAM、ZRAM、Flash）。在一個輕量級的「物件年齡演算法」驅動下，該控制器動態地將物件分類為熱、溫、冷三種狀態，並在無需依賴預測模型的情況下，將其遷移至最佳的儲存層級。此架構與 Android 執行環境的垃圾回收機制深度整合，採用了「類書籤式」機制，以確保記憶體正確性，並避免昂貴的喚回成本。 我們在 Android 10 環境中的全面評估證明了此方法的有效性。實驗結果顯示，PhantomSwap 顯著提升了應用程式的快取容量，在商業應用測試中能同時支援多達 19 個應用程式，而原生 Android 系統僅能支援 15 個。至關重要的是，PhantomSwap 在維持極具競爭力的中位數熱啟動效能的同時，展現了卓越的「尾部延遲」表現，在複雜的遊戲負載下，其第 95 百分位的啟動延遲較頂尖學術框架降低了達 46%。其中的 ZRAM 層作為關鍵緩衝區，有效減緩了記憶體未命中時的效能衝擊，並防止了雙層系統中常見的嚴重停頓問題。 總而言之，本研究證明了 PhantomSwap 的反應式三層式架構，為提升記憶體容量並在行動裝置上提供更一致、更具韌性的使用者體驗，提供了一個穩健且實用的解決方案。

Mobile devices face a persistent conflict between the growing memory demands of sophisticated applications and their limited physical DRAM. This conflict leads to aggressive process termination by the Low Memory Killer (LMK), which degrades the multitasking user experience. While prior object-level swapping frameworks like Marvin and Fleet have sought to address this by optimizing for single performance vectors, such as reclamation speed or hot-launch latency, they can exhibit performance cliffs under complex, unpredictable workloads. This thesis introduces PhantomSwap, a hybrid memory swapping system for mobile devices designed to deliver a more holistic improvement in system fluidity and resilience. PhantomSwap's core innovation is a reactive, three-tiered memory hierarchy (DRAM, ZRAM, Flash), managed by an intelligent Unified Swap Controller. Driven by a lightweight Object Aging Algorithm, the controller dynamically classifies objects as hot, warm, or cold, and migrates them to the optimal storage tier without relying on predictive models. This architecture is deeply co-designed with the Android Runtime's Garbage Collector, employing a Bookmark-style mechanism to ensure memory correctness without expensive fault-ins. Our comprehensive evaluation on a reproduced Android 10 environment demonstrates the effectiveness of this approach. Experimental results show that PhantomSwap significantly increases application cache capacity, supporting up to 19 concurrent commercial applications compared to 15 in the Android baseline. Critically, while maintaining highly competitive median hot-launch performance, PhantomSwap exhibits superior tail latency, reducing the 95th-percentile launch time by up to 46% against state-of-the-art frameworks in complex gaming workloads. The ZRAM tier acts as a crucial buffer, "softening" the performance penalty of memory misses and preventing the severe stalls inherent in two-tiered systems. In conclusion, PhantomSwap provides a robust and practical solution that enhances system capacity and delivers a more consistent, resilient user experience on mobile devices.