基於多通道固態硬碟減少峰值電流下快閃記憶體管理方法

Abstract

近年來，固態硬碟為了效能提升而採用多通道平行架構，加上固態硬碟容量增加，固態硬碟峰值電流問題也逐漸被討論，但多是討論省電議題，峰值電流研究較少被提及。當固態硬碟內所有晶片一起運作時，峰值電流量相當驚人。但是傳輸介面能提供的電流預算有限，加上特定條件下系統有可能會限制儲存裝置的電流預算。當可用預算減少時晶片將無法全數運作，效能勢必受到影響，如何讓供電受限制下效能退化幅度減少為本研究最重要之目的。此外，為了能控制峰值電流，我們首先會建立峰值電流模型以獲得峰值電流資訊，並用此資訊配合電流預算去管理工作晶片選擇，在建立模型過程我們發現每個元件都會有基本耗電，我們利用這個特性推出峰值電流模型公式。另外我們提出了兩個運作晶片受限制時會讓效能退化的緩衝區與通道競爭問題，並發現兩者問題本質上存在衝突關係，因此我們提出Adaptive Contention Management (ACM)去做動態調整。實驗結果發現使用ACM方法可以比隨機挑選改善吞吐量20%以上。

In recent years, SSDs enhance the overall performance with adopting a multi-channel architecture. With SSDs capacity and performance increase, more and more reasearchs discuss its power consumption issue. However, most of reasearchers talk about power saving problem, only a few papers concern with peak current topic. The peak current of chips in SSDs are incredible when they program in a same time. Besides, there are not enough power budget of host interface for SSDs. Chips can’t program simultaneously with insufficient current, it caused SSDs performance down seriously. So, how to control and minimize performance degrade is our most important challenge. To understand SSDs peak current information, we would build peak current model first, and we found foundament current in each flash parallelism component, so we can used this to construct peak current model. After model was completed, we can use information of model to determine which chips could be actived. There had two trade-off contention issue would effect performance behavior, called buffer space contention and channel time contention, we proposed Adaptive Contention Management(ACM), a dynamic manage policy to migitage contention problem. Experiments shows that overall performance better than Rand_B policy of at most 20% in a variety of environments.