Author : Chin-Hsien Wu Presenter : kilroy
Outline Introduction Related work Motivation Main idea Evaluation Conclusion Q & A
Introduction Flash memory is a popular alternative for the design of storage systems for lots of reasons: Shock-resistant Energy-efficient Non-volatile nature This thesis focus on NAND flash because of : NAND flash is more suitable for designing file/storage systems Dr. Fujio Masuoka
Related work Flash-memory characteristics Write-once characteristic Size of Page/block Cumbrous overhead Previous Work Garbage collection Wear leaving Energy consumption
Motivation B-Tree is one of the popular data structures adopted by database systems However, a direct adoption of B-Tree could exaggerate the overhead of flash management
Main idea The design and implementation of BFTL:
Main idea (cont.) Reservation Buffer: Dirty records should be timely flushed to flash -> “index unit” Index unit consists of : Original B-Tree component 、 identifier and op_flag Note that BFTL use FTL to store index units Index unit will pack into a logical term “sector”
Main idea (cont.) The Commit Policy Index units of the same node could be scattered over sectors
Main idea (cont.) Node translation table: Sectors will scatter over flash memory
Evaluation Experiment setup and performance metrics 4 MB NAND flash Reservation buffer size : 60 record The bound of the lengths of lists in the node translation table : 4
Evaluation(cont.) Performance of B-Tree Index structures creation The workloads consisted of insertions only Inserted records
Evaluation(cont.) Performance of B-Tree Index structures maintenance
Evaluation(cont.) Performance of B-Tree Index structures search
Evaluation(cont.) The size of reservation buffer and energy consumption issues
Evaluation(cont.) The size of reservation buffer and energy consumption issues
Conclusion Original B-Tree not suitable over flash memory However, a BFTL implement doing well on flash memory