本研究提出 Autopilot，一套應用於 Fault-Tolerant Virtual Machine （FTVM）之 Runtime Adaptive Checkpoint Optimization 機制，並實作於 NCU-MFTVM 平台。相較於既有方法多依賴固定 Epoch、Threshold 或 Heuristic Rule 進行調整，Autopilot 以 dirty page rate 作為 runtime observation signal，動態反映 workload intensity、synchronization overhead 與 client latency，並根據系統狀態自動調整 Epoch Length。本研究進一步建立 Runtime Adaptive Control Framework，包含 Idle、Correction 與 Hold 三種控制狀態，以持續搜尋並維持接近最佳之 Epoch Length，降低 synchronization overhead 與 client latency。實驗採用 Apache Benchmark、TPC-C 與 Rsync workload，並於不同 Host Spec 與 VM configuration 下進行評估。實驗結果顯示，Autopilot 可根據 workload 與系統環境動態調整 Epoch Length，在 latency-oriented、throughput-oriented 與 network bandwidth-oriented workload 下皆能有效提升 application performance 與 runtime stability，並於不同硬體環境中維持穩定且接近最佳之 epoch length。;This study proposes Autopilot， a runtime adaptive checkpoint optimization mechanism for Fault-Tolerant Virtual Machines (FTVM)， implemented on the NCU-MFTVM platform. Unlike existing methods that rely on fixed epochs， thresholds， or heuristic rules， Autopilot utilizes dirty page rate as a runtime observation signal to dynamically reflect workload intensity， synchronization overhead， and client latency， and automatically adjusts epoch length according to application workload. To achieve stable runtime optimization， this study further designs a runtime adaptive control framework consisting of Idle， Correction， and Hold states， enabling the system to continuously search for and maintain near-optimal epoch lengths while reducing synchronization overhead and client latency. Experiments using Apache Benchmark， TPC-C， and Rsync workloads under different host specifications and VM configurations demonstrate that Autopilot can dynamically adapt epoch length to varying workloads and system environments， effectively improving application performance and runtime stability across latency-oriented， throughput-oriented， and network bandwidth-oriented workloads.
