Experimental evaluation of behavior-based failure-detection schemes in real-time communication networks

Effective detection of failures is essential for reliable communication services. Traditionally, non-real-time computer networks have relied on behavior-based techniques for detecting communication failures. That is, each node uses heartbeats to detect the failure of its neighbors and the end-to-end transport protocol (e.g., TCP) achieves reliable communication by acknowledgment/retransmission. Recently, there has been a growing demand for reliable `real-time' communication, but little research has been done on the failure detection problem. In this paper, we present two behavior-based failure-detection schemes - neighbor detection and end-to-end detection - for reliable real-time communication services and experimentally evaluate their effectiveness. Specifically, we measure and analyze the coverage and latency of these detection schemes through fault-injection experiments. The experimental results have shown that nearly all failures can be detected very quickly by the neighbor detection scheme, while the end-to-end detection scheme uncovers the remaining failures with larger detection latencies.