Research And Implementation Of Object-based Distributed Coordination Services System

Many large-scale distributed systems depend on coordination services,an infrastructure software that provides a replication and synchronization framework for distributed applications.The coordination services need to replicate data on different machines for liveness,and thus they also need to gurantee the replications consistency for safty.It is impossible to guarantee safety and liveness in an unreliable distributed system at the same time.In order to handle this problem,Chubby and Zookeeper use Paxos and ZAB,the most powerful existing consistnecy protocals,respectively to provide guaranteed safety,and also ensure a very high level of liveness.In this paper,we implement a complete distributed coordination service system Sabber.Sabber is based on a novel Raft replicated state machine implementation that employs an object-oriented approach in which the system actively creates and maintains live replicated objects for user-provided objects.Compared to the file-based Zookeeper approach,this object-based approach avoids separating control-flow from data-flow and it is easier and more natural to implement coordination service.When clients access the remote objects using remote procedure calls(RPCs),Sabber guarantees the linearizable semantics to clients.Sabber supports a variety of method calls,which makes clients access these remoted objects as local ones,and thus makes it more flexible and efficient for users to implement.In order to get a better performance,we modify the original Raft protocol on several aspects.Sabber also includes other modules such as log compaction,snapshot,dynamically reconfiguring,client interaction to act as a comprehensive system.Experiments show that Sabber can guarantee safety.As for the write and read ability,Sabber is comparable to Zookeeper.We also show that Sabber has higher throughput than Zookeeper in coordination services such as distributed lock,distributed queue,distributed counter.