Optimization in sustain random write performance for enterprise SSD

Abstract

Thanks to the advances in NAND flash memory technology, Solid State Drives (SSDs) are prevailing in storage device markets. Though SSDs are succeeding in enterprise markets, random write performance is one of the major concerns in enterprise SSD. Because enterprise workloads most likely require random writes. In addition, SSD is weak at random write workloads. Considering that SSD using the DRAM optimized scheme such as demand map loading, the random write operation in SSD meets several obstacles for its performance. Such as garbage collection and map load / unload operations. These operations occupy the channels of NAND, which limits the performance. So in this thesis, we introduce new architecture and its management to lower the occupation of the NAND channels and optimize the random write performance. We achieved this by adding SLC-only NAND dies and separating the map data and user data operation physically by dividing the channel. We compared the results with die-added parallelized base architecture and base architecture in our emulator. And found improvement in random write performance.