Evaluation of message delay correlation in distributed systems

Partially synchronous computational models fall in between synchronous models, which are demanding in terms of requirements but admit solutions to most of the important fault-tolerant distributed computing problems, and the purely asynchronous model, where this is not the case. The ⊗-Model is a recently presented partially synchronous model close to pure asynchrony. It just assumes a bound on the ratio of maximum vs. minimum end-to-end delays of messages simultaneously in transit. This paper provides experimental evidence for the assumed correlation of end-to-end delays in some common type of distributed systems: using a comprehensive custom evaluation framework, we measured the end-to-end delays in a simple clock synchronization algorithm running on a fast-Ethernet network connecting Linux workstations. Our results reveal a significant correlation between maximum and minimum delay under several load conditions, and hence confirm that the bounded ratio assumption of the ⊗-Model is well-founded.