Data Management Strategy In Mixed Ramcloud Architecture

RAMCloud,the new storage architecture that was born in the era of big data,has been developed in recent years due to its advantages of fast access,easy expansion and large storage capacity.RAMCloud stores data in dynamic random access memory?DRAM?at any time,thereby eliminates the disadvantages of high latency and low throughput of Disk access,and using memory replace Disk,Data access and processing can be performed more quickly.Due to its volatile DRAM cannot save data for a long time,RAMCloud guarantees the data durability by backing up several copies of data from the Master server to the backup server,thereby increasing the storage cost,and using memory to store data makes the system's total energy increases.Phase-change memory?PCM?is a new type of storage media that emerged in recent years.Due to its non-volatile,high storage density and low energy consumption.it has taken the place of DRAM in the vision of academia and industry.However,PCM also has two major weaknesses:its limited lifetime compared with DRAM whose lifetime is 10¹⁵;PCM's read and write speeds are not equal,writing speed is much slower than reading speed.Thus It is impossible to use PCM completely replace DRAM as memory to storage data.We must consider combining the advantages of the two devices and optimize the known shortcomings to build the mixed RAMCloud architecture.This thesis studies the RAMCloud architecture and optimize its shortcomings,using PCM replace part of DRAM built a same-level mixed RAMCloud architecture.Taking into account the PCM's life is far less than the DRAM's,if the write count and wear leveling control are not performed,the system will automatically crashed when PCM memory expires,this thesis discusses the probability of data whether exist in the PCM within specified period of time,if the probability of the existence of data is higher than 0.7,no longer write it into PCM but directly scan the PCM memory.Simultaneously build a doubly linked list,by dynamically counting the access count of the page in the PCM,always keep the list monotonically increasing from tail to head,and ensure that data always written at the end of the list.Considering data object management,Firstly,the memory allocation ratio of the mixed RAMCloud is calculated by comparing the read energy consumption,write energy consumption,standby energy consumption,and active energy consumption of the two devices,and it is found that the system storage performance is optimum when the ratio of the DRAM and the PCM is 4.847:1.Later prejudge the read/write tendency of the data,store the data with high read tendency in the PCM,and store the rest data in the DRAM.This step is to reduce the amount of PCM writing data.In order to ensure the stability of the architecture,build a data migration model,manage an LRU linked list in each PCM master,and dynamically migrate data object with low read tendency in PCM to DRAM,at the same time use cost-benefit analysis to delete the data which with low importance.The mixed RAMCloud architecture builds good performance in data reading,power-off recovery,guaranteed hit rate and data storage.