Parallel garbage collection in Solid State Drives

Flash memories are making their way into both desktop and server environments. Over the years, the major limitation to the wide-adoption of flash memories has been their cost. However, with the advancements in the semiconductor industry, the price per gigabyte (GB) gap between the conventional disk drives and flash memories is getting closer. As such, flash memories can replace disks, where disk utilization is less and extra spindles are added just to increase performance. Though they ventured into the storage architecture as cache and as a hybrid counterpart with Hard Disk Drives (HDD), slowly they are expected to replace the disk drives in servers and super computers [1]. The other major drawback with flash memory is its inability to sustain unlimited erase cycles, which directly limits their lifetime [2]. In order to improve reliability, it is proposed to create redundancy [3].;Creating a Redundant Array of Independent Disks (RAID) is a conventional way of providing redundancy in hard disk drives (HDD) [4]. The same idea is adopted in Solid State Drives (SSD). In addition to the conventional RAID techniques that are implemented at the device level (external RAID), redundancy can be created in an SSD at a much lower level (internal RAID) [3]. The scope of this work is limited to internal RAID.;This work uses i-RAID [3]; an architecture and simulator for internal RAID as background and proposes two improvements. The first contribution is to improve the dynamic stripe formation using access patterns. Another enhancement is to utilize the idle domains when i-RAID is not active by invoking parallel instances of garbage collection. This thesis describes how these methods can affect the performance of the device and explains how the internal parallelization of an SSD can be better exploited. Both the methods are evaluated individually and the findings are presented. Though both the methods have a great potential to improve the performance of the device, the earlier work (on which the current work is based) is done in such a way that exploiting access patterns during stripe formation could not provide much improvement.;[1--4] Please see thesis for references.