End-to-end latency of a fault-tolerant CORBA infrastructure

This paper presents measured probability density functions (pdfs) for the end-to-end latency, of two-way, remote method invocations from a CORBA client to a replicated CORBA server in a fault-tolerance infrastructure. The infrastructure uses a multicast group-communication protocol based on a logical token-passing ring imposed on a single local-area network. The measurements show that the peaks of the pd/s for the latency are affected by the presence of duplicate messages for active replication, and by the position of the primary server replica on the ring for semi-active and passive replication. Because a node cannot broadcast a user message until it receives the token, up to two complete token rotations can contribute to the end-to-end latency seen by the client for synchronous remote method invocations, depending on the server processing time and the interval between two consecutive client invocations. For semi-active and passive replication, careful placement of the primary server replica is necessary to alleviate this broadcast delay to achieve the best possible end-to-end latency. The client invocation patterns and the server processing time must be considered together to determine the most favorable position for the primary replica. Assuming that an effective sending-side duplicate suppression mechanism is implemented, active replication can be more advantageous than semi-active and passive replication because all replicas compete for sending and, therefore, the replica at the most favorable position will have the opportunity to send first.