Research On SSD Cache Optimization Based On Data Block Graph

With the advancement of non-volatile memory,flash-based SSD has been widely used.SSD read and write performance is between DRAM and HDD,and the cost is relatively low compared to DRAM,so it is usually deployed as a large cache between DRAM and HDD.However,the traditional caching algorithm mainly focuses on the increase of hit rate,ignores the amount of data written to cache,which will greatly shorten the lifetime of SSD due to that SSD can sustain only a finite number of erase/write cycles.To address this challenge,We analyzed the actual workload and found that there are many write operations that are unnecessary,such as low access probability data or periodic backup data.For traditional caching algorithms,this data will still enter the cache and be replaced continuously.If we can prevent this data from entering the cache in advance,it can not only keep the hot blocks in the cache for a longer period of time but also eliminate unnecessary SSD writes and avoid cache pollution.So we propose a cache architecture based on block graph analysis named CacheRank.In this architecture,we use the sliding window technology to mine the relationship between blocks and analyze the block sequence to construct a data block graph(block graph).The vertices in the graph represent data blocks,the edges represent the access relationships between the data blocks,and the Dice coefficients are used to quantify the strength of the association between the blocks and save them in the attributes of the edges.Then we propose a BlockRank algorithm suitable for block importance analysis based on the block graph and use the calculated block importance as the cache admission indicator of the data block.Experiments show that the architecture is effective in reducing SSD cache writes.In the case of LRU,the write traffic is reduced by 25%-41%,and the hit rate is increased by 1%-4%.