Optimization of pre-read to improve performance of DRAMless SSD

Abstract

In this thesis, a new pre-read method to improve the performance of DRAMless SSD is presented. DRAMless SSD has cost competitiveness, but due to limitations in memory buffer size, it can quickly reach a bottleneck of host I/O, resulting in relative degradation in average performance inevitable. Due to the characteristics of hardware feature of DRAMless SSD like no DRAM and small SRAM size, its performance degradation is inevitable. However, to achieve a purpose to improve the value of DRAMless SSD by reducing the gap in performance with Conventional SSD, dynamic pre-read was proposed. Dynamic pre-read can optimize buffer utilization by processing pre-read for sequential I/O in parallel and diminish impact of pre-read fail. As a result of the experiment, performance of dynamic pre-read improved by about 16% to 22% compared to when pre-read was not applied and improved by about 6.6% to 9.4% compared to when fixed pre-read is applied, in sequential read workload. Also, Performance improved by about 9.5% compared to when pre-read was not applied and by about 4.3% compared to when fixed pre-read is applied, in mixed workload.