Analysis of Clustering Page Allocation

NAND-based Solid-State Drives (SSDs) have become the new generation of storage devices for modern computer systems. SSDs employ Page Allocation techniques to deal with a large volume of requests from applications. CFDP and PFCD static allocation schemes are most widely used as they allow write I/O requests to be handled in parallel. However, as the streaming platforms characterized by a high volume of long read traces have become more and more popular, the traditional page allocation schemes have revealed their limitations on improving read performance due to the dispersed allocation of related data. In addition, CFDP and PFCD page allocation schemes have suffered from the trade-off between the utilization of system-level parallelism and internal flash-chip features, which hinders SSDs from achieving higher read performance. This thesis proposed the Clustering Page Allocation (CPA) scheme to improve read performance by utilizing advanced commands and at the same time exploiting parallel access to different channels. The trade-off between those features is optimized with the cluster threshold by considering the size of application-level I/Os and the number of channels/flash-chips to maximize the read performance. Compared to CFDP and PFCD, the CPA scheme achieves a higher throughput with a factor of 1.14 - 2.56 and a lower latency with a factor of 0.39 - 0.87.

MAJOR ADVISOR: Ben Lee
COMMITTEE: Thinh Nguyen
COMMITTEE: Bechir Hamdaoui
COMMITTEE: Lizhong Chen
GCR: Jimmy Yang