利用建構在MRAM上的緩衝快取增進I/O效能

Abstract

MRAM(Magnetic Random Access Memory)是最近這幾年來新興的一項記憶體技術，其具有非揮發性的特性，並且有著跟傳統DRAM相似的存取速度，而且一般研究認為MRAM是短期間最可能被量產且取代現有DRAM的技術。而由於傳統緩衝快取的設計是建立在非揮發記憶體上，因此若是將非揮發性記憶體取代揮發性主記憶體，傳統的緩衝快取僅能藉由延長將資料寫回的時間以增進系統效能。因此我們設計一個架構在MRAM 上的緩衝快取，考慮到除了利用記憶體非揮發性的特性延長修改過的資料寫回磁碟的時間，也可以作其他的改進以增進效能。因此在MRAM-based Buffer Cache的設計中，我們提出了兩個主要的機制: (1)在VFS和一般檔案系統之間加入一層 ramfs 檔案系統。其主要的目的是暫時儲存新建立的檔案，延遲檔案分配給其真正所屬的檔案系統的時間，等到較為確定檔案大小後再進行分配，以避免檔案 fragmentation的情形並減少不必要的I/O；(2)修改以往以檔案為單位將資料寫回的演算法，加入了考慮每個要寫回的block在磁碟上相對應的位置，讓較連續且較鄰近的資料一起寫回，減少寫回所需的時間

MRAM(Magnetic Random Access Memory) is an emerging technology in recent years with non-volatile characteristic and access speed comparable to DRAM(Dynamic Random Access Memory). It is generally believed that MRAM may completely replace DRAM and begin mass production in the short-term. Traditional buffer cache design is based on volatile memory, so after MRAM completely replace DRAM as system main memory, we can only take advantage of delaying write to improve system performance. Therefore we purpose a MRAM-based buffer cache considering that not only exploit non-volatile characteristic to propagate new data to disk, but also exploit non-volatile system main memory to further improve I/O efficiency. In MRAM-based buffer cache architecture, we purpose two main policy:(1) Adding ramfs filesystem between VFS and general filesystem. The main purpose is to store files temporarily ,so we can delay the time to allocate file to its original filesystem until file size is more certain to avoid file fragmentation. (2) Purpose a writeback algorithm, considering file spatial locality, to flush continuous and near blocks together to reduce writeback time.