Research And Optimization Of Wear-aware Load Balancing For SSD-based Distributed Storage System

With the development of cloud computing,the performance demands on the storage are increasing.However the traditional disk-based storage system cannot meet the demand.Due to faster access performance and lower power consumption,SSDs are widely used to accelerate data-intensive applications.However,the existing load balancing scheme is designed for the traditional disk,and it does not consider the asymmetry of read and write costs of SSD,and a limited lifetime.The existing wear leveling solutions based on distributed file systems are mostly based on custom SSDs or SSD simulators,and their versatility is not strong.For all-SSD-based distributed file system,we have designed a wear-aware load balancing scheme called Wear-Aware Load Balancing(WALB).We designed a wear model based on the SSD usage and the amount of data written by the host.WALB performs data balance by sensing the read load of the data server and the wear value of the SSD.When the wear of the cluster is not balanced,the hot update data,hot reading data,and cold data on the heavier node will be migrated to the node with light wear.When the load of read is not balanced,the hot reading data will be migrated from the heavier node to the node with lighter load.When wear imbalance and read load imbalance occur at the same time,the priority is given to wear leveling.For newly created files,the client of Cappella will caches them.When the client has cached a lot of data items,the data will be flushed to the lightly-worn data nodes.The experimental results show that compared with the load balancing strategy,WALB can increase IOPS by 13% to 74.5%,and reduce the amount of migration data by 19.2% to 34%,and compared with the EDM strategy,increase IOPS by 17.5% ~ 31.7%,in the write-intensive workload,reduce the amount of migration data by 10.8% to 16.3%.Compared with the EDM strategy,the wear leveling scheme has a relatively more uniform distribution,and the wear leveling strategy reduces the number of writes by 1.6% to 7.8%,and reduces the number of erasures by 4.1% to 10.8%.