A Paxon Evaluation of P2

P2 is a new and exciting system that permits declarative implementations of overlay networks. The recently published work cites reasonable implementations and performance for two protocols, Chord and Narada. Our work evaluates P2 through an implementation of a third protocol, the Paxos consensus protocol. P2’s declarative language Overlog, admitted a smooth and readable implementation of the Paxos preliminary protocol [1]. Even though this implementation was a correct declaration of the Paxos preliminary protocol, P2’s lack of concurrency control and well defined execution semantics provided a source difficult bugs and race conditions, and we remain uncertain as to wether or not our implementation is robust. Our initial conclusions are that Overlog is good language for implementing declarative overlays, but that P2 must provide more clear semantics and concurrency mechanisms for the system to become a widely adopted overlay network platform. To evaluate the performance of P2, we built a simulator that could replay the P2 messages for various sized Paxos consensus instances, providing a lower bound on the time to reach consensus. We compared running times and CPU % utilization for identical conversations generated by executing P2 and by running our simulator called p2replay. Our preliminary experiments showed that P2 performed linearly with respect to the number of nodes, but was significantly slower and consumed much more of the CPU than the p2replay execution. For 256 nodes, P2 performed about 10 times as slow and consumed roughly 4 times as much of the CPU.