Research On Space Collection Schemes For Erasure-coded Clustered In-memory Stores

In recent years,in order to improve memory space efficiency while ensuring data availability,erasure codes have gradually been applied to data-intensive memory clusters to store data.When reclaiming memory space,there are hot and cold data blocks in the stripes at the same time.As a result,the cold data blocks cannot be eliminated from the memory in time,resulting in low memory space utilization and affecting system response performace.To address this problem,this paper studies the space collection scheme for erasure-coded clustered in-memory stores.In order to improve memory space utilization and system access performance,a correlation-aware memory space collection scheme(CaMSC)is proposed.CaMSC designed a replacement update strategy(RUS)to improve memory space utilization.RUS first completes the reclaiming of some cold data blocks through stripe reorganization,and then RUS replaces the cold data blocks in the stripes that have not been reorganized with newly entered memory data blocks to ensure all cold data blocks can be eliminated to improve memory space utilization.However,the RUS strategy generates additional replacement and update traffic when reclaiming cold data blocks,resulting in low efficiency of memory reclaiming.So a correlation-aware stripe organization strategy(CSOS)is designed to reduce the times of replacing cold data blocks,thereby reducing the replacement and update traffic and speeding up the reclaiming of cold data blocks.CSOS analyzes the associated data blocks in a data access stream through frequent sequence mining algorithms,and organizes the correleated data blocks into same stripes,thereby increasing the probability that multiple data blocks in the same stripe will be eliminated at the same time,and reducing the times of replacement;In addition,CSOS has also optimized the original data mining algorithm to improve the efficiency of the data mining algorithm by reducing the times of iterations of the original data mining algorithm and reducing the generation of invalid frequent sequences.Both the CaMSC scheme and two candidate space collection scheme(i.e.,temporallocality-aware memory space collection scheme(Ta MSC)and basic memory space collection scheme(BMSC))are implemented within a real-world erasure-coded in-memory store,where the traces generated by YCSB benchmark are replayed to quantitatively evaluate the above three schemes.The results of experiment show that when the two comparison schemes do not implement the RUS strategy,compared with Ta MSC and BMSC,CaMSC improves memory space collection rates by 86.5% and 97.75%,and reduces the average access delay by 15.1% and 17.8%,respectively.When the two comparison schemes implement the RUS strategy,compared with Ta MSC and BMSC,CaMSC reduces the replacement and update traffic of 13.2%?17.4% and 45.3%?50.9%,respectively,and CaMSC reduces 10%?17.4% and 34.4%? 37.8% of the memory space collection time.