Font Size: a A A

Mpaxos:Low Latency And High Throughput Consensus Under Geographically Shifting Workloads

Posted on:2021-01-25Degree:MasterType:Thesis
Country:ChinaCandidate:Y C WangFull Text:PDF
GTID:2428330620968180Subject:Software engineering
Abstract/Summary:PDF Full Text Request
With the rapid development of the Internet,there are more and more users of Web services,and they come from all over the world.Many global web services have to respond to changes in clients' access characteristics to achieve higher throughput and lower latency.To achieve high throughput and low latency,it is a common practice to deploy multiple local web services that only serve closing clients.However,multiple different services mean higher management and maintenance costs,and it's difficult to achieve real-time data exchange.So it cannot support systems that require global high availability and consistency,such as international shopping websites and international financial transaction systems.Therefore,it is a very challenging task for these systems to be able to adapt to changes in workload characteristics.The main contributions of this paper are summarized as follows:1.Proposed MPaxos based on leaderless Paxos protocol and introduced the concept of Working Cluster.The Working Cluster is the set of nodes currently participating in the voting,and it is a subset of the MPaxos cluster.The number of nodes in the working cluster W satisfies W = 2F + 1.In this way,by introducing a working cluster,F can be decoupled from N(the total number of MPaxos nodes),allowing more nodes to be deployed around the world without affecting the efficiency of the protocol.2.A Reorganization process for adjusting the working cluster is proposed.The reorga-nization process is decentralized while avoiding the occurrence of live locks,which makes the migration process very fast and minimizes the impact on throughput and delay.3.Separate the working clusters used by different data,each data object has a working cluster,so each object is allowed to adjust its working cluster to optimal according to its load characteristics.4.A scheduling framework suitable for MPaxos is proposed,which can calculate the optimal working cluster members under a given load.We implemented MPaxos in Paxi,an open source consensus protocol framework.And we proved that MPaxos performs well in a WAN environment through a large number of experiments.
Keywords/Search Tags:Consensus Protocol, Global Web Service, Shifting Workload, Migration
PDF Full Text Request
Related items