緩和在多通道下自我治癒結構之固態硬碟的治癒風暴問題

Abstract

為了降低每個位元的價位，快閃記憶體的儲存密度越來越高，這會造成快閃記憶體的壽命越來越低，而這也是快閃記憶體要能普及所需關注的重要議題，近來的研究指出損壞的區塊可以藉由自我治癒來增加快閃記憶體的壽命，然而傳統磨損平均被用來平衡所有區塊的抹擦次數，也就是說所有幾乎所有區塊會在同一時間自我治癒，我們稱這集中治癒的問題為治癒風暴，治癒風暴會造成較高的反應時間、較差的穩定性還有突然急遽升高的耗電量，得益於傳統磨損平均可以平衡區塊抹擦次數的特性，我們在每個區塊加上一個擁有不同初始值的虛擬抹擦次數，並且讓傳統的磨損平均作用在上面，藉由如此我們可以讓所有區塊的抹擦次數分佈在一個可控的範圍內並藉此分散區塊治癒的時機，我們還有提出一個方法用較早的去做區塊治癒來避免讀請求會被區塊治癒所卡住，並且我們的方法可以以幾乎沒有額外負擔的情況下做在現有的快閃記憶體轉換層和磨損平均演算法上，在我們的實驗中我們測試了許多不同的測資，我們成功的呈現了我們的方法在沒有影響快閃記憶體壽命的前提下解決了治癒風暴的問題。

The flash storage density continues to increase because of high cost per bit. However, it causes worse flash endurance and short life time is a critical issue of the popularity of NAND flash memory. Recent studies show that worn-out block can be healed by heat-accelerated self-recovery to extend flash life time. However, the erase counts of all blocks are balanced under conventional wear leveling, and thus all blocks would heal almost at the same time. We call intensive block healing as a heal storm. A heal storm may result in bad response time and throughput, poor reliability, and instantaneously high power consumption. Inspired by evening distribution of erase counts under conventional wear leveling, we add a virtual erase count to each block with a different initial value and let conventional wear leveling operate on it. In this way, the distribution of erase counts would be distributed in a control interval and therefore disperse the healing time. We further propose a method to do block healing earlier to avoid read requests blocked by heal commands. Moreover, our approach can apply on existing FTL design and wear leveling algorithm with almost no extra overhead. In our experiments, we successfully show that our approach solves the heal storm problem without affecting the flash life time under different kinds of workloads.