Seventh International Conference on Simulation Tools and Techniques

Research Article

Parallelism Potentials in Distributed Simulations of Kademlia-Based Peer-to-Peer Networks

Download35 downloads
  • @INPROCEEDINGS{10.4108/icst.simutools.2014.254609,
        author={Philipp Andelfinger and Konrad J\'{y}nemann and Hannes Hartenstein},
        title={Parallelism Potentials in Distributed Simulations of Kademlia-Based Peer-to-Peer Networks},
        proceedings={Seventh International Conference on Simulation Tools and Techniques},
        publisher={ICST},
        proceedings_a={SIMUTOOLS},
        year={2014},
        month={8},
        keywords={network simulation distributed kademlia dht lookahead},
        doi={10.4108/icst.simutools.2014.254609}
    }
    
  • Philipp Andelfinger
    Konrad Jünemann
    Hannes Hartenstein
    Year: 2014
    Parallelism Potentials in Distributed Simulations of Kademlia-Based Peer-to-Peer Networks
    SIMUTOOLS
    ICST
    DOI: 10.4108/icst.simutools.2014.254609
Philipp Andelfinger1,*, Konrad Jünemann1, Hannes Hartenstein1
  • 1: Karlsruhe Institute of Technology
*Contact email: philipp.andelfinger@kit.edu

Abstract

The benefits of distributing a network simulation depend on characteristics of the simulated network. Performance improvements reported in the literature are comparatively low for peer-to-peer overlay networks in particular, as the logical topology of these networks can necessitate frequent synchronization between the processors executing the simulation. In this paper, we show that a speedup of up to a factor of 6.0 using 16 nodes connected using InfiniBand and close to linear reductions in memory usage are possible for simulations of Kademlia-based networks. Our distributed simulator implementation enables simulations of one of the largest peer-to-peer networks at full scale of about 10 million peers. Based on the two fundamental goals of minimizing communication between processors and minimizing synchronization overheads, we propose two strategies for assigning simulated nodes to processors. We analyze the effects of the two strategies and show that each strategy supports one of the goals, while being detrimental to the other. We propose efficiency metrics that expose how much of the potential for parallel execution is exploited by a simulator. Through detailed performance measurements and by applying the new metrics to our simulator implementation, we quantify remaining efficiency potentials.