應用於固態硬碟上的高效能寫入緩衝管理演算法

Abstract

SSD使用NAND Flash Memory為基本的儲存元件具有讀取速度快、抗振性佳等優點，在部份的應用上逐漸取代傳統硬碟成為系統的主要儲存裝置。但因為Flash Memory的物理特性為在寫入資料前必需先進行抹除的動作，由於抹除單位比寫入單位大很多的緣故，所以需要對random write的寫入動作特別處理，否則會導致寫入效能不佳成為系統的效能瓶頸。而且因為檔案系統內的資料破碎以及Multi-Thread的應用，寫到儲存裝置的write pattern會更加random，如何改善因為random write造成的SSD寫入效能下降問題愈顯重要。我們提出一個應用在SSD上的新Buffer管理演算法HitStat，除了同時考慮時間區域性和空間區域性外，並能夠動態的調整得到較佳的Padding設定來大幅改善寫到NFTL的write pattern。最後實驗結果顯示，HitStat與既有SSD上的緩衝管理方法FAB、BPLRU相比，寫入效能平均改善了30%以上。

The base storage component of SSD is NAND Flash Memory, it has some advantages such as the high speed of read operation and shock resistance. On some application areas, SSD gradually replaces traditional hard drive disk becoming the major storage device of entire system. But the physical property of Flash Memory is Erase-before-Write, because of the erase-unit is much larger than the write-unit, we need take care of the random write pattern on SSD, otherwise the speed of write operation will degradation and become the bottle neck of entire system. Furthermore, due to disk fragmentation and applications of Multi-Thread, the write pattern will become more random, how to improve the problem of write throughput degradation on SSD caused by random write becoming more and more important. We proposed a new buffer management algorithm applied to SSD called HitStat, not only consider temporal locality and spatial locality simultaneously, but also adjust the Padding Threshold to better settings dynamically, much improve the write pattern written to NFTL. Finally, it shows about 30% enhancement of write throughput compared to FAB and BPLRU, which are other buffer management algorithms on SSD.