Hybrid Distributed Consensus

Inspired by the proliferation of cloud-based services, this paper studies consensus, one of the most fundamental distributed computing problems, in a hybrid model of computation. In this model, processes (or nodes) exchange information by passing messages or by accessing a reliable and highly-available register hosted in the cloud. The paper presents a formal definition of the model and problem, and studies performance tradeoffs related to using such a register. Specifically, it proves a lower bound on the number of register accesses in deterministic protocols, and gives a simple deterministic protocol that meets this bound when the register is compare-and-swap (CAS). In addition, two efficient protocols are presented; the first one is probabilistic and solves consensus with a single CAS register access in expectation, while the second one is deterministic and requires a single CAS register access when some favorable network conditions occur. A benefit of those protocols is that they can ensure both liveness and safety, and only their efficiency is affected by the probabilistic and timing assumptions.