Who's On Board? Probabilistic Membership for Real-Time Distributed Control Systems (Extended)
To increase their dependability, distributed control systems (DCSs) need to agree in real time about which hosts have crashed, i.e., they need a real-time membership service. In this paper, we prove that such a service cannot be implemented deterministically if, besides host crashes, communication can also fail. We define implementable probabilistic variants of member- ship properties, which constitute what we call a synchronous membership service (SYMS). We present an algorithm, ViewSnoop, that implements SYMS with high-probability. We implement, deploy and evaluate ViewSnoop analytically as well as experimentally, within an industrial DCS framework. We show that ViewSnoop significantly improves the dependability of DCSs compared to membership schemes based on classic heart- beats, at low additional cost. Moreover, ViewSnoop distinguishes, with high probability, host crashes from message losses, enabling DCSs to counteract losses better than existing approaches.