The Investigation And Implementation Of Optimized File System Technologies For Solid State Disks

With the development of computer system, the performance gap between external storage devices and CPU is increasingly larger. Hard disk driver (HDD) as mainstream storage is the bottleneck to improve overall system. Comparing to HDD, the flash-based solid state disk (SDD) has the advantages of low read delay, fast random read and write, low power consumption and high reliability. Moreover, the manufacturing process of SSD will be more mature and I/O performance will be improving. Eventually, it is advised that SSD will replace the current HDD in the future.The SSD`s technical characteristics is different from HDD, and traditional file system design and optimization is on the assumption that storage system consists of HDDs, so the file system cannot fully play the advantages of SSD. An investigation and implementation of optimized file system technologies for SSD is made in the paper which focuses on how to take full advantage of the SSD. What we do lays the foundation for high-performance file system for SSD.The research is divided into two levels. First, it is from the view of file system buffer cache. SSD has asymmetric write/read operation. The overhead of a write operation is many times higher than that of a read operation. Hence, we must keep acceptable hit ratio and delay evicting the dirty pages as late as possible to let down the write miss ratio. The page replacement algorithm for SSD aims to reduce the write counts so that it can enhance the overall I/O performance of storage system. We suggest a novel algorithm called AML (Adaptively Mixed List) which prefers to evict the cold clean page and changes the window size adaptively according the history of I/O access pattern. Finally, we conduct an overall trace-driven simulation. The result shows AML reduces the runtime on average by 14% compared to CFLRU.Second, the attention is paid to the file system technologies themselves including block allocation scheme, allocation delay of requested space, write delay, read ahead and so on. Due to SSD's high read performance and overwriting before erase, the optimization of random-write performance is the most important for it. Through the analysis of basic block allocation scheme in the traditional file system, we propose the flash-aware block allocation scheme called FBA (Flash-aware Block Allocation). FBA classifies files by size, and takes different approach according the classification. Besides, ideal file system block size is involved in the FBA. By reducing the file system fragmentation, FBA decrease the write overhead caused by the merger. Along with space allocation delay and write-delay optimization of hidden delay technology, random-write performance of file system is enhanced a lot. Since the SSD's read latency which can be hidden is very small and read-ahead mechanism has a limited impact on system performance, we make an optimization on it. Eventually, we achieve these optimized technologies in the traditional file system and conduct tests to demonstrate higher performance of the file system after optimization. The performance of random write improves at most by 11%.