| In order to maintain the data consistency of replicas in the distributed system,some well-known solutions are proposed,such as Paxos,Raft,and their variants.The existing consistency protocols usually adopt the strong leadership approach,in which the leader sorts commands and distributes messages.However,with the increase of the number of replicas and the expansion of the system scale,the leader will gradually become the bottleneck of the system performance.In addition,today’s service clusters are usually deployed across multiple WANs.Long distance transmission will inevitably lead to greater communication delay.Access in various regions also shows the characteristics of spatial locality.However,Paxos and other protocols are more suitable for deployment in small-scale clusters,which is difficult to adapt to WAN scenarios.Therefore,how to optimize Paxos protocol in WAN scenarios and reduce its leader bottleneck is a very challenging topic.Aiming at the above problems,this thesis proposes the following solutions to optimize the throughput and delay of the system and reduce the bottleneck of leaders:(1)A distributed consistency protocol GPaxos based on replication group and relay is proposed.The task of sending and collecting messages from the leader is handed over to the relay node for processing,which reduces the leader’s load pressure and alleviates the leader’s bottleneck problem.At the same time,the group is divided according to the geographical distribution characteristics of nodes.The relay in the group is responsible for communicating with the leader,and the leader can collect messages from multiple groups.Overlapping between groups is allowed,which increases the link redundancy of communication.Each round of relays will be selected according to the priority list,and the priority list will also be updated to select the node with the least delay and the best performance to undertake the relay task.The timeout of GPaxos will be dynamically adjusted according to the delay,so it has better adaptability in the case of network fluctuations.The experimental evaluation shows that GPaxos can significantly improve throughput and latency in the scenario of multi node cross regional distribution.(2)A leader migration algorithm based on multiple leaders is proposed.Multiple leaders are allowed to exist in the system at the same time.Each leader has the leadership of some objects,and each adjacent region has the leader of hot data in the region,so that the local region can submit data faster.At the same time,the leader migration algorithm is introduced to allow the leadership of objects to change between different regions when hotspot access transfers occur between regions.Through experimental verification,the design based on multi-leader can replace the leader in time when hot access transfer occurs,and the maximum throughput is increased by about 25%. |
