Seventh International Conference on Simulation Tools and Techniques

Research Article

Large-Scale Network Simulation: Leveraging the Strengths of Modern SMP-based Compute Clusters

Download527 downloads
  • @INPROCEEDINGS{10.4108/icst.simutools.2014.254622,
        author={Mirko Stoffers and Sascha Schmerling and Georg Kunz and James Gross and Klaus Wehrle},
        title={Large-Scale Network Simulation: Leveraging the Strengths of Modern SMP-based Compute Clusters},
        proceedings={Seventh International Conference on Simulation Tools and Techniques},
        publisher={ICST},
        proceedings_a={SIMUTOOLS},
        year={2014},
        month={8},
        keywords={parallel simulation shared-memory distributed simulation},
        doi={10.4108/icst.simutools.2014.254622}
    }
    
  • Mirko Stoffers
    Sascha Schmerling
    Georg Kunz
    James Gross
    Klaus Wehrle
    Year: 2014
    Large-Scale Network Simulation: Leveraging the Strengths of Modern SMP-based Compute Clusters
    SIMUTOOLS
    ICST
    DOI: 10.4108/icst.simutools.2014.254622
Mirko Stoffers1,*, Sascha Schmerling1, Georg Kunz1, James Gross2, Klaus Wehrle1
  • 1: RWTH Aachen University
  • 2: KTH Royal Institute of Technology
*Contact email: stoffers@comsys.rwth-aachen.de

Abstract

Parallelization is crucial for efficient execution of large-scale network simulation. Today's computing clusters commonly used for that purpose are built from a large amount of multi-processor machines. The traditional approach to utilize all CPU cores in such a system is to partition the network and distribute the partitions to the cores. This, however, does not incorporate the presence of shared memory into the design, such that messages between partitions on the same computing node have to be serialized and synchronization becomes more complex. In this paper, we present an approach that combines the shared-memory parallelization scheme Horizon [9] with the standard approach to distributed simulation to leverage the strengths of today's computing clusters. To further reduce the synchronization overhead, we introduce a novel synchronization algorithm that takes domain knowledge into account to reduce the number of synchronization points. In a case study with a UMTS LTE model, we show that both contributions combined enable much higher scalability achieving almost linear speedup when simulating 1,536 LTE cells on 1,536 CPU cores.