雙重混合快閃記憶體轉換層:運用寫入分配以及切割預留空間在頁級轉換快閃記憶體層與混合轉換快閃記憶體層之策略

Abstract

固態硬碟使用快閃記憶體為儲存元件具有輕巧、耐震且儲存密度高等優點，在應用上逐漸取代傳統硬碟成為一種新的趨勢。目前的一般使用電腦越來越發達，且程式多支援多線程，使得寫到儲存裝置的寫入模式越來越凌亂，因此，如何改善該狀況下造成固態硬碟寫入效能下降問題便顯得相當重要。本研究提出一個應用在固態硬碟上的新架構，將頁級轉換快閃記憶體層與混合轉換快閃記憶體層做結合，搭配寫入群分配，將寫出的資料群分派到合適的快閃記憶體層。另外，由預留區塊調整分割給各快閃記憶體轉換層。實驗結果顯示，資料群資料量大的適合寫入混合轉換快閃記憶體層，資料群資料量小的適合頁級轉換快閃記憶體層，且在調整預備區塊後，能得到一個最佳預備區塊分配點，使得整體回收成本達到最低。且當垃圾回收壓力增加，效能更加顯著。

Real-life workload is mixed with long/sequential and small/random bursts, however, to optimize these two patterns in the single FTL makes design become more complicated. In this paper, we proposed a new design which takes advantages of hybrid mapping FTL and page mapping FTL to handle sequential and random pattern respectively. We focus on extracting different types of write patterns from the workload and dispatching appropriate write groups into each suitable FTL. Moreover, we apply overprovision partitioning mechanism for lower overall flash management cost because garbage collection overhead is inversely proportional to its overprovision size. Experimental result shows there are existing optimal partition point in each workload, and significantly reduce write amplification while increasing garbage collection pressure.