Performance Optimization Of Log-structured File System Based On Hotness

NAND Flash has the advantages of high performance and low latency,so it is widely used in storage systems at all levels.Log-structured File System(LFS)is very friendly to Flash features.It transforms random writes into sequential writes by out-of-place updates,with higher random write performance and lower consistency maintenance overhead.However,LFS still has problems such as excessive segment cleaning overhead and serious write amplification from file systems,which leads to a significant reduction in the performance of LFS and the lifespan of Flash device.The common data hotness difference in I/O workload aggravates the problems of segment cleaning and write amplification.Aiming at those problem in LFS,an optimization method of log-structured file system based on hotness is proposed,which is called OLH.OLH adopts hot and cold data identification strategy based on data clustering,and dynamically identifies hot and cold data by sampling and clustering the hotness metadata.During the writing process,hot and cold data are written into different logs respectively to reduce the segment cleaning overhead.For hot data,OLH defers writing it to the log according to the hotness value to reduce the write amplification from "wandering tree" caused by its frequent updates.At the same time,in order to avoid the performance degradation caused by random write introduced by multiple logs,OLH also defers submitting logs to the file system to achieve the purpose of batch writing data to improve write performance.During the segment cleaning process,a hotness-based victim selection algorithm is used to fully consider the influence of hotness on the victim,and when the victim is released,only the dirty node data related to the victim is flushed,to reduce the segment cleaning overhead and the the write amplification from consistency maintenance.The hotness metadata caching strategy is designed and implemented to reduce the overhead of accessing hotness metadata introduced by OLH.We verify the effectiveness and performance of OLH using several realistic workloads.Experimental results show that OLH can adapt to changes in hotness.Compared with F2 FS optimization strategy,segment cleaning overhead is reduced by36.3%-36.5%,bandwidth is increased by 62.8% to 71.4%.In terms of write amplification from file systems,the total write volume of file and node data decreased by 6.8% to 9.3%,and the total write volume of file system metadata decreased by44.4% to 53.1%.In terms of the overhead introduced,the hit rate of the hotness metadata cache can reach 64.3% to 97.3%.Most of the hotness metadata reading can be done in memory,and the access overhead is small.