Maintaining Determinism in Replicated Middleware Applications

摘要

Non determinism can lead to data inconsistency state in actively replicated environment. This defeats the purpose of replication as a fault-tolerance strategy. We proposed an efficient method which ensures determinism among the replicas in fault tolerance middleware applications. This method exploits the technique of statically analysing the application source code of client, and finds the variables and system calls which leads to non determinism in the replicas. A flag field is introduced to indicate that the source code of the client consists of non-deterministic variables, and it is sent to the servers together with client request. Flag is verified in servers, initiates the distributed coordination method if it is true else sends any one response of the servers to the client by duplicate removal. In distributed coordination method, all replicas execute incoming requests; one replica is elected as primary based on the time stamp value. It is responsible for taking all non?? deterministic decisions. The state of the primary replica is updated to all other replicas connected asynchronously to maintain consistency. This method reduces the communication overhead. We evaluate our technique for the active replication of servers using micro benchmarks that contain various sources of non-determinism. i.e. multi-threading, system call, shared I/O and random ( ).