Fault-tolerant TCP mechanisms

Abstract

While fault-tolerance is supported by a variety of critical services that can be accessed over the Internet, they are not robust in that they are oblivious of the impact of their tolerant mechanisms on the service they deliver. Throughputs and fail-over latencies are the most suitable metrics for a fault-tolerant service. We propose a few fault tolerant TCP mechanisms to improve the overall throughput and provide efficient failure detection and recovery in the existing HYDRANET-FT infrastructure. Synchronizing the receiving TCP states of the server replicas, primary and backup, using TCP reassembly queue improves the overall throughput. TCP retransmissions and replica management daemons are used for failure detection and fail-over. Recovery of a host after failure involves dynamic rebuilding of TCP state to that of the existing hosts in the server group. Measurements on an experimental testbed show that our mechanisms result in a robust fault-tolerant implementation with a substantial improvement in the overall throughput for a single primary-backup system.