Impact of WAN channel behavior on end-to-end latency of replication protocols

Software replication of stateful services is typically implemented using two-tier architectures, in which clients directly interact with replicas running distributed agreement protocols for ensuring consistency. In general, performance of these protocols is sensitive to network delays, which might consequently reduce service availability. Therefore, in previous works we introduced three-tier software replication, in which agreement protocols run in an apposite tier (detached from clients and replicas) that can be independently deployed in a controlled and stable part of the network. In this paper, we analyze the performance of replication protocols implemented using two- and three-tier architectures using a simplified wide-area network model that considers two types of behaviors for channels, i.e., normal (small and predictable latency variations), and slow (high and unpredictable latency variations). This channel model is instantiated using traces of real Internet latencies measured sending HTTP requests to Internet web-sites at varying rates. Then, by exploiting traces, we simulate simplified versions of three replication protocols (i.e., active, passive, and three-tier replication), and we show how the end-to-end latency of each protocol is related to the number of slow channels. Results mainly demonstrate that the availability of a service replicated through a three-tier architecture is less affected from channel slow-downs.