固態硬碟通道管理演算法之讀取修改寫入相依性

SSD Channel Management with Read-Modify-Write Dependencies

陳冠群(Guan-Chun Chen)

指導教授 ： 張立平

國立交通大學/資訊學院/資訊科學與工程研究所/碩士(2014年)

https://doi.org/10.6842/NCTU.2014.00913

摘要

近幾年的固態硬碟，都會利用多通道架構增加其效能，為了使其多通道可以完整發揮其平行寫入效果，因此會加入寫入緩衝區，以收集由主控端來之寫入要求，收集到一定量之後再寫入各通道，可避免通道因無寫入要求而閒置，寫入通道與邏輯頁之間的綁定採用動態綁定，意即邏輯頁可以由寫入緩衝區指派到隨意的通道上，在上述的條件下，由於主控端的寫入單位以及固態硬碟的寫入單位不一致，因此由主控端來的寫入要求會有個嚴重的問題，便是寫入要求寫不滿一個固態硬碟之實體頁面，這樣會導致必須做讀取修改寫入之行為，此行為本身就是一種額外成本，會拉低效能，在頁面越來越大的近年趨勢，此問題會越加嚴重，此外讀取修改寫入行為在動態綁定的情況下會產生額外閒置時間，之前文獻提出解決閒置時間之方法，卻會使得產生額外成本，本篇論文提出透過預先讀取，導致不會因為資料相依性而產生額外閒置時間，也因為只預先讀取而非改變逐出順序，因此不產生其他不必要額外成本。實驗在各種工作環境下勝過其他方法可達8%至20%。

關鍵字

固態硬碟 ； 寫入緩衝 ； 讀取修改寫入

並列摘要

In recent years, SSDs employ a multi-channel architecture to enhance the overall performance,in order to make all channels can be used evenly,write buffer is added to collect write requests so that channel can avoid become idle state. Besides ,binding between the channel and the logical page using dynamic binding, which means that the logical page can assigned by the write buffer to any channel ,under the above conditions, due to the host write unit and SSD write unit inconsistent ,so there is a serious problem, when write request can’t satisfied entity page ,this will cause the behavior named read-modify-write ,this behavior itself is an overhead, it would lower overall performance , in the page size growing trend in recent year, this problem become increasingly serious.Besides,read-modify-write will generate additional overhead in the case of dynamic binding.Prior paper purpose methods to resolve idle time, but it will incur additional costs.So we propose Lazy merge,through this methods ,lower idle time generated , and because the evict order is not changed , thus no other additional costs.Experiments shows that overall performance better than other methods of 8-20% in a variety of environments.

並列關鍵字

Nand flash ； Write buffer ； Read-modify-write

參考文獻

Technical Conference, pages 57–70, 2008.

design for high-performance flash disks. ACM SIGOPS

Operating Systems Review, 41(2):88–93, 2007.

algorithm exploiting multi-chip parallelism in solid state disks.

[4] Dongwook Kim, Sooyong Kang. Partial Page Buffering for Consumer Devices with