Optimization Of Cloud Storage Qo S Based On Deep Reinforcement Learning

With the rapid growth of the global data volume,human beings have entered the era of big data,and the first problem facing them is how to store these massive data.Because the traditional centralized data storage methods have many restrictions,such as high cost and poor scalability.It can no longer meet the storage requirements of massive data.Therefore,various distributed storage systems have emerged at the historic moment.The QoS performance of the system has become a research hotspot.An excellent data distribution strategy is the key to designing a distributed storage system,which determines how to distribute massive data to data storage nodes as evenly as possible,and the balance of data distribution has a great impact on data read and write performance.Affects the QoS performance of the cloud storage system.CRUSH algorithm,as the core algorithm of data distribution in Ceph distributed storage system,is a pseudorandom data distribution algorithm with high efficiency and high scalability.However,the CRUSH algorithm has the problem that the data is unevenly distributed on the storage nodes and affects the read and write QoS performance,which is specifically manifested in two aspects: storage nodes with a large amount of data will inevitably bear the load of heavy reads and writes,and it is easy to become a system performance bottle;on the other hand,the relatively low utilization of storage nodes with a small amount of data,is a waste of resources and reduces the QoS performance of the system.Aiming at the above problems,the work of this paper is based on the current widely studied and used cloud storage system Ceph.Its main work and innovation are as follows:(1)In a small-scale scenario,aiming at the problem that the Ceph data distribution algorithm CRUSH has an uneven distribution of data on the device space,which ultimately affects the read and write QoS performance of the cloud storage system,this paper designs and implements a data based on reinforcement learning The distribution algorithm RLCRUSH,first performs theoretical analysis and experimental verification from the data distribution process and the algorithm itself.It is concluded that the uneven distribution of PGs on the OSD nodes is the cause of the uneven distribution of user data on various storage devices.The three steps of modeling,designing a training model based on reinforcement learning,and designing and implementing a new model-based algorithm are used to improve the existing CRUSH algorithm,so that the PG can be approximately evenly distributed to each OSD node,eliminating uneven data distribution.The resulting system bottleneck,which improves OSD disk usage and QoS performance of cloud storage systems.(2)For medium and large-scale scenarios,the environment becomes more complicated due to the increase in scale.We have designed and implemented a new data distribution algorithm DRL-CRUSH based on deep reinforcement learning.The ability to quickly extract and transform information combined with the ability of reinforcement learning to make automatic decisions and control improves the original data distribution algorithm.The experimental results show that the algorithm still has obvious distribution equilibrium of PG among OSD nodes in medium and large-scale scenarios.Improvement,and has better time performance.(3)At the same time,the current algorithms based on reinforcement learning are mostly used in video games,robot control,and unmanned driving.We try to introduce them into the cloud storage field,and use reinforcement learning to continuously interact with the environment,feedback and automatic decision-making.And the ability to optimize goals to optimize the QoS performance of cloud storage systems,and hope to bring some inspiration to researchers in different fields.