Title: Fault-Tolerant Distributed Protocols in Ordered Networks
Advisors: Dan Ports and Tom Anderson
Abstract: Protocols for fault-tolerant distributed systems typically assume a completely asynchronous network, in which messages can be dropped or reordered arbitrarily. Guaranteeing the safety of client-server based distributed storage systems in this context involves guaranteeing two distinct properties: order of state updates and their durability. This work shows that in single datacenters, separating these two concerns and guaranteeing ordering in the network using the capabilities of upcoming programmable switches can avoid most of the latency and throughput costs of replication and consistency. This work studies two related problems---state machine replication and distributed transaction processing---and presents two protocols---NOPaxos and Eris---which leverage different network ordering properties to deliver drastically improved normal case performance.