快閃儲存系統上之資料配置研究

Abstract

快閃記憶體（Flash Memory）已被廣泛使用於嵌入式系統各種可攜帶是電子配備。隨著技術的精進，快閃記憶體的價格不斷下降、容量也不斷增加，因此也逐漸開始應用在各式大型儲存系統上，如預備用於硬碟的固態硬碟(Solid State Disk)。作為儲存設備的元件，效能和壽命是很重要的。在效能方面，因為快閃記憶體沒有任何需要移動的元件，固態硬碟常使用的增進效能方式是將多顆快閃記憶體晶片透過匯流排的連接、讓其可以同時平行運作；此組織方式常稱為多晶片(Multi-chip)或多儲存庫(Multi-bank)快閃儲存系統。在壽命方面，快閃式記憶體的壽命受限於它的清除次數。在多晶片快閃儲存系統上，不同的資料配置方式，會影響快閃記憶體清除次數。 此論文即在分析不同的資料配置方式配合不同的晶片組織架構所造成的清除次數與效能效果。在清除次數方面，由於用戶端(Host)給予的一個要求內的資料之間常有相關性(locality)，因此將屬於同一個要求的資料寫在一起，可以增加清除效能並降低清除次數。而在效能方面，如果於系統不忙碌的狀態下，盡量將資料分散寫在不同晶片上，可以提高效能；但是當系統非常忙碌的情況下，將資料集中寫入和分散寫入的效能非常相近，反而集中寫入減少清除動作，還有可能使效能增加。 我們建立了一個使用模擬追縱導向(Trace-driven)方式的模擬器，模擬快閃儲存系統的運作，藉此觀察不同的資料配置在不同的晶片組織架構下的效果。

In the field of embedded systems and consumer electronics, Flash memory has emerged as an excellent storage system. As prices continue to decline and capacities increase, NAND Flash SSD is showing the potential to substitute Hard Disk Drives. As a secondary storage, the reliability and performance issues are important. The lifetime of Flash-based SSD is limited by the erase count of Flash blocks. Under the multi-package Flash organization, different data allocation policies can lead to different number of erase operations. In this thesis, the effect on performance and erase count resulting from different data allocation policies and Flash architectures is studied. In terms of erase count, preserving data locality by keeping request un-striped leads to less erase count. In terms of performance, the experimental results show that if the workload is light, stripe request to different banks effectively reduces response time. However, when the workload is relatively heavy to the serving Flash architecture, keeping request un-striped still has comparable performance or even better by the assistance of erase count reduction.