A performance vs. cost framework for evaluating DHT design tradeoffs under churn

Protocols for distributed hash tables (DHTs) incorporate features to achieve low latency for lookup requests in the face of churn, continuous changes in membership. These protocol features can include a directed identifier space, parallel lookups, pro-active flooding of membership changes, and stabilization protocols for maintaining accurate routing. In addition, DHT protocols have parameters that can be tuned to achieve different tradeoffs between lookup latency and communication cost due to maintenance traffic. The relative importance of the features and parameters is not well understood, because most previous work evaluates protocols on static networks. This paper presents a performance versus cost framework (PVC) that allows designers to compare the effects of different protocol features and parameter values. PVC views a protocol as consuming a certain amount of network bandwidth in order to achieve a certain lookup latency, and helps reveal the efficiency with which protocols use additional network resources to improve latency. To demonstrate the value of PVC, this paper simulates Chord, Kademlia, Kelips, OneHop, and Tapestry under different workloads and uses PVC to understand which features are more important under churn. PVC analysis shows that the key to efficiently using additional bandwidth is for a protocol to adjust its routing table size. It also shows that routing table stabilization is wasteful and can be replaced with opportunistic learning through normal lookup traffic. These insights combined demonstrate that PVC is a valuable tool for DHT designers.