Previous editions‎ > ‎PADABS 2015‎ > ‎

Program 2015

9:00-10:30 First Session: Welcome to PADABS and Panel "Distributed Open Agent-Based Simulation Benchmark"
  • Welcome (Vittorio Scarano, Gennaro Cordasco and Ugo Erra)
  • Panel: Distributed Open Agent-Based Simulation Benchmark (Chair Paul Richmond), results and discussions available on http://www.openab.org.
11:00-12:30 First Session: Load Balancing on Agent-Based Simulations (Chair Vittorio Scarano)

  • Nick Collier, Jonathan Ozik and Charles Macal, “Large-Scale Agent-based Modeling with Repast HPC: A Case Study in Parallelizing an Agent-based Model”. 
    • Abstract: We present a case study for parallelizing a large-scale epidemiologic ABM developed with Repast HPC. The original serial model is a CA-MRSA model, which tracks CA-MRSA transmission dynamics and infection in Chicago, and represents the spread of CA-MRSA in the population of Chicago. We utilize both within compute node parallelization using the OpenMP toolkit and distributed parallelization across multiple processes using MPI. The combined approach yields a 1350% increase in run time performance utilizing 128 compute nodes. pdf
  • Alessia Antelmi, Gennaro Cordasco, Carmine Spagnuolo and Luca Vicidomini, “On Evaluating Graph Partitioning Algorithms for Distributed Agent Based Models on Networks”. 
    • Abstract: Graph Partitioning is a key challenge problem with application in many scientific and technological fields. The problem is very well studied with a rich literature and is known to be NP-hard. Several heuristic solutions, which follow diverse approaches, have been proposed, they are based on different initial assumptions that make them difficult to compare. An analytical comparison was performed based on an Implementation Challenge (David A. Bader et al. 2012), however being a multi-objective problem (two opposing goals are for instance load balancing and edge-cut size), the results are difficult to compare and it is hard to foreseen what can be the impact of one solution, instead of another, in a real scenario. In this paper we analyze the problem in a real context: the development of a distributed agent-based simulation models on a network field (which for instance can model social interactions). We present an extensive evaluation of the most efficient and effective solutions for the balanced k-way partitioning problem. We evaluate several strategies both analytically and on a real distributed simulation settings (D-MASON). Results show that, a good partitioning strategy strongly influence the performances of the distributed simulation environment. Moreover, we show that there is a strong correlation between the edge-cut size and the real performances. Analyzing the results in details, we were also able to discover the parameters that need to be optimized for best performances on networks in ABMs. pdf
  • Claudio Márquez, Eduardo Cesar and Joan Sorribes, “Graph-Based Automatic Dynamic Load Balancing for HPC Agent-Based Simulations”. 
    • Abstract: The main problem of Agent-Based Modelling (ABM) simulations in High Performance Computing (HPC) is load imbalance due to a bad distribution of the agents that may generate uneven computation and increase communication overhead, inhibiting the efficiency of the available computing resources. Moreover, the agents’ behaviours can considerably modify the workload at each simulation step thereby affecting the workload progression of the simulation. In order to mitigate such problems, automatic mechanisms for dynamically adjusting the computation and/or communication workload are needed. For this reason, we introduce an Automatic Dynamic Load Balancing (ADLB) strategy to reduce imbalance problems as the simulation proceeds. The ADLB tunes the global simulation workload migrating groups of agents among the processes according to their computation workload and their message connectivity map modelled using a Hypergraph. This Hypergraph is partitioned using the Zoltan Parallel HyperGraph partitioner method (PHG). In addition, to prevent excessive all-to-all communications, the ADLB uses filtering routines to send message groups to specified recipient processes in a simple 3D grid-based structure. Our method has been tested with a biological ABM using the framework Flexible Large-scale Agent Modelling Environment (Flame), obtaining a significant impact on the application performance. pdf
14:00-16:00 Second Session: Parallel and Distributed Agent-Based Simulations  (Chair Carmine Spagnuolo)

  • Biagio Cosenza, “Behavioral Spherical Harmonics for Long-Range Agents’ Interaction”.
    • Abstract: We introduce behavioral spherical harmonic (BSH), a novel approach to efficiently and compactly represent the directional-dependent behavior of agent. BSH is based on spherical harmonics to project the directional information of a group of multiple agents to a vector of few coefficients; thus, BSH drastically reduces the complexity of the directional evaluation, as it requires only few agent-group interactions instead of multiple agent-agent ones. We show how the BSH model can efficiently model intricate behaviors such as long-range collision avoidance, reaching interactive performance and avoiding agent congestion on challenging multi-groups scenarios. Furthermore, we demonstrate how both the innate parallelism and the compact coefficient representation of the BSH model are well suited for GPU architectures, showing performance analysis of our OpenCL implementation. pdf
  • Peter Heywood, Paul Richmond and Steve Maddock, “Road Network Simulation using FLAME GPU”. 
    • Abstract: Demand for high performance road network simulation is increasing due to the need for improved traffic management to cope with the globally increasing number of road vehicles and the poor capacity utilisation of existing infrastructure.This paper demonstrates FLAME GPU as a suitable Agent Based Simulation environment for road network simulations, capable of coping with the increasing demands on road network simulation. Gipps' car following model is implemented and used to demonstrate the performance of simulation as the problem size is scaled. The performance of message communication techniques has been evaluated to give insight into the impact of runtime generated data structures to improve agent communication performance. A custom visualisation is demonstrated for FLAME GPU simulations and the techniques used are described. pdf
  • Alban Rousset, Bénédicte Herrmann, Christophe Lang and Laurent Philippe, “A communication schema for parallel and distributed Multi-Agent Systems based on MPI”. 
    • Abstract: The interest for Multi-Agents Systems (MAS) grows rapidly and especially in order to simulate and observe large and complex systems. Centralized machines does not however offer enough capacity to simulate the large models and parallel clusters can overcome these limits. Nevertheless, the use of parallel clusters implies constraints such as mono-threaded process of execution, reproductibility or coherency. In this paper, our contribution is a MPI based communication schema for Parallel and Distributed MAS (PDMAS) that fits High Performance Computing (HPC) on cluster requirements. Our communication schema thus integrates agent migration between processes and it guarantees message delivery in case of agent migration. pdf
16:30-18:00 Third Session: Distributed Agent-Based Simulations and Practice (Chair Gennaro Cordasco)

  • Davide Cingolani, Alessandro Pellegrini and Francesco Quaglia, “RAMSES: Reversibility-based Agent Modeling and Simulation Environment with Speculation-support”.
    • Abstract: Techniques for speeding up the execution of complex agent-based models, such as Parallel Discrete Event Simulation (PDES), are known to be of great relevance/benefit. In this paper, we present RAMSES, a framework for easily specifying agent-based discrete event models entailing both environment and agent entities, which offers parallel execution capabilities based on speculative event processing---recognized as a core support for scalability of the PDES engine---and an innovative {\em software reversibility} technique---to cope with state restore in case the run slides along a non-consistent speculative path. Software reversibility in RAMSES relies on automatic static software instrumentation, thus allowing the model developer to concentrate on the actual forward-execution logic of the events occurring in the system. In fact, RAMSES transparently generates at runtime update undo code blocks which are able to cancel back the effects of the execution of events, relieving the modeler from the burden of the (error-prone) implementation of the state recoverability support enabling speculative processing. An experimental assessment of RAMSES is additionally presented, when comparing it to highly-optimized general-purpose PDES-based simulation frameworks. pdf
  •  Nicola Lettieri, Carmine Spagnuolo and Luca Vicidomini, “Distributed Agent-based Simulation and GIS: An Experiment With the dynamics of Social Norms”.
    • Abstract: In the last decade, the investigation of the social complexity has witnessed the rise of Computational Social Science, a research paradigm that heavily relies upon data and computation to foster our understanding of social phenomena. In this field, a key role is played by the explanatory and predictive power of agent-based social simulations that are showing to take advantage of GIS, higher number of agents and real data. We therefore focus on the issues that need to be considered for distributing GIS based ABMs. We observed that the density distribution of agents, over the field, strongly impact on the overall performances. In order to better understand this issue, we analyzes three different scenarios ranging from real positioning, where the citizen are positioned according to the dataset to a random positioning where the agent are positioned uniformly at random on the field. Results confirm our hypothesis and show that an irregular distribution of the agents over the field increases the communication overhead. We provide also an analytic analysis which, in a 2-dimensional uniform field partitioning, is function of several parameters (which depend on the model), but is also influenced by the density distribution of agents over the field. According to the presented results, we have that uniform space partitioning strategy does not scale on GIS based ABM characterized by an irregular distribution of agents. pdf
  •  Bilge Acun, Nikhil Jain, Abhinav Bhatele, Misbah Mubarak, Christopher D. Carothers and Laxmikant Kale, “Preliminary Evaluation of a Parallel Trace Replay Tool for HPC Network Simulations”. 
    • Abstract: This paper presents a preliminary evaluation of TraceR, a trace replay tool built upon the ROSS-based CODES simulation framework. TraceR can be used for predicting network performance and understanding network behavior by simulating messaging on interconnection networks. It addresses two major shortcomings in current network simulators. First, it enables fast and scalable simulations of large-scale supercomputer networks. Second, it can simulate production HPC applications using BigSim’s emulation framework. In addition to introducing TraceR, this paper studies the impact of input parameters on simulation performance. We also compare TraceR with other network simulators such as SST and BigSim, and demonstrate its scalability using various case studies. pdf