應用於固態硬碟之跨層協同式效能最佳化策略

Abstract

基於快閃記憶體的固態硬碟，目前已經廣泛地被使用於筆記電腦，智慧手機，與平板電腦。固態硬碟有著非常獨特的內部資料管理機制以及效能最佳化手段。在傳統的架構下，作業系統與儲存裝置兩者，僅能透過固有的區塊裝置介面交換讀寫命令以及資料。然而，一方面作業系統端無從得知儲存裝置內部的資料管理狀況，二方面裝置端喪失了檔案系統層次的資訊。以上種種問題，限制了改善固態儲存裝置效能的潛在機會。 本研究探討如何強化傳統的區塊裝置介面，從而使作業系統端與儲存裝置端可以互相交換資訊，藉跨層的協同手段，達成最佳化固態儲存裝置效能的目的。此外，亦可透過強化的區塊裝置介面，使得主機端可取得裝置內部的管理資訊，從而以協助的角度來進行耗費較多計算時間或者空間成本的演算方法。本計劃提案的技術議題，則為基於此強化的溝通介面下，探討什麼樣的工作適合跨層次協同處理，以及協同處理的相關策略。本計劃預計以三年為期，第一年討論跨層的緩衝快取協同管理，亦即作業系統端分頁快取可查詢裝置內部快取以及垃圾收集狀況，決定將哪些資料優先往下傳遞給裝置快取。第二年探討快閃記憶體轉址的協同處理，也就是說主機端從裝置取得部份的轉址資訊，並視機會進行高解析度轉址，減輕裝置端進行轉址時造成的效能與空間成本。第三年預計討論管理演算法所需的計算資源之分攤，亦即主機端從裝置端獲得管理資訊後，可執行較複雜且空間需求較大的管理演算法，如垃圾收集以及冷熱資料辨認等等，並將計算結果交給裝置來執行。我們相信藉由資訊以及資源共享的方式，主機端的作業系統與裝置端的軔體能協同發掘出更大的效能躍進。

Flash-based storage devices have been widely used in laptops, smartphones, and tablets. Flash memory has very different physical characteristics, imposing unique challenges on its data management. Traditionally, the host OS and storage devices exchanges read-write commands and data via the standard block-device interface. However, beyond this line the storage device losses file-system-level information and conversely, the host OS cannot know how data is managed inside of the storage device. This design sacrifices performance to layered structures. In this research project, we shall investigate how to extend the traditional block-device interface such that the host OS and the flash-storage device can exchange useful information for not only collaborative management and computational-resource sharing. This project is planned to be 3-year. In the first year, we shall investigate possible collaboration between the page cache in the host and the write buffer in the device. The page cache can query the storage device about the current status of buffer management and garbage collection and then makes a better decision on which data should be pushed down to the device level. In the second year we will explore cooperative address translation between the host and the device. The host can retrieve mapping information from the device and relieve the device of the overhead related to high-resolution mapping. The work of the third year aims at computation-resource sharing: The host can carry out algorithms of garbage collection and hot-data identification with high time and space complexity based on necessary information retrieved from the device. We believe such cross-layer collaboration can unleash new potential of performance improvement.