Improving random write performance of NAND flash-based solid state drives

Abstract

The NAND flash-based solid state drive~(SSD) has emerged as revolutionary storage media thanks to its attractive properties over hard disk drives~(HDDs).However, the wide-spread of SSDs is impeded by the poor random write performance of SSDs.In spite of advances in the SSD technology, the random write performance is still lower than the sequential write performance of SSDs by a factor of 10.This thesis proposes two approaches to remedy this problem.The first approach is to transform harmful random writes into more favorable write patterns outside SSDs.To conduct this transformation over an SSD, we developed a software layer, called \textit{ReSSD}, which works as a virtual block device on top of the SSD.By inspecting all incoming requests, ReSSD identifies small random writes which have potential to degrade SSD`s performance significantly.ReSSD writes the data of the identified small random writes sequentially to the reserved area of the SSD and eventually moves them to their original locations in an ordered-sequential write fashion.This approach is effective in improving the performance of \textit{low-end SSDs} whose random write performance is only comparable to or lower than that of HDDs due to stringent cost constraint.The second approach is to improve the efficiency of garbage collection inside SSDs by exploiting the information of precise data liveness.We devised a new flash management technique, called \textit{Zombie Chasing}, considering dirty data in the buffer cache.Zombie Chasing introduces a new liveness state, called the \textit{zombie} state, to denote the live data that will be dead shortly due to the corresponding dirty data in the buffer cache.Zombie Chasing distinguishes such zombie data from the real live data to optimize garbage collection algorithms.The zombie-aware garbage collection utilizes the zombie data information when selecting a victim block and treats the zombie data as the hot data that will be overwritten soon.We implemented the...