RFLRU: A Buffer Cache Management Algorithm for Solid State Drive to Improve the Write Performance on Mixed Workload

摘要

Flash memory based Solid State Drives (SSD)acquiring greater attention in enterprise storage computingenvironment; this is primarily due to its high I/O speed. SSDsuse multiple NAND flash memory chips as a storage media anddeploy internal RAM to maintain the flash translation layer(FTL) mapping table. The rest portion of the inner RAM isused as a buffer. This buffer absorbs the write requests andthus the resulting write requests to the NAND flash memory isdetermined by the SSD buffer cache management algorithm.Many of the previously proposed write buffer cachemanagement algorithms concentrate on improving the randomwrite performance either by reordering the writes, addressingthe temporal locality or evicting the clean pages etc. And theyhave not concentrated enough to exploit the sequential localityin the write pattern. Because of this, the input sequential writepatterns are not completely utilized by the Flash TranslationLayer and that reduces the number of switch merges andincreases full and partial merges in log block based FTL. Inthis paper, a novel algorithm called RFLRU: Random FirstLeast Recently Used is proposed to improve the performanceof SSD write operation on mixed workload. The algorithmidentifies the interleaved sequential writes, constructs variouscache eviction policies and the write sequence is constructed bycontemplating the SSD characteristics. Essentially this newtechnique reduces number of erase and writes operations onSSD. An efficient RFLRU implementation is developed andtested in a trace driven simulation environment and comparedto the previously proposed LRU FAST, BPLRU and REFbuffer management schemes. The results show that RFLRUreduces the number of merge; erase and write operations andthe overall write performance is improved.