Large scale multiagent simulation on the grid

Lecture Notes in Computer Science, 2009

We describe a distributed architecture for situated largescale agent-based simulations with predominately local interactions. The approach, implemented in AglobeX Simulation platform, is based on a spatially partitioned simulated virtual environment and allocating a dedicated processing core to the environment simulation within each partition. In combination with dynamic load-balancing, such partitioning enables virtually unlimited scalability of the simulation platform. The approach has been used to extend the AgentFly air-traffic testbed to support simulation of a complete civilian air-traffic touching National AirSpace of United States. Thorough evaluation of the system has been performed, confirming that it can scale up to a very high number of complex agents operating simultaneously (thousands of aircraft) and determining the impact of different configurations of the simulation architecture on its overall performance.

Advances in Intelligent and Soft Computing, 2011

This paper presents the work done during the first year of the EPIS project. This project deals with the process of conducting multiple and parallel multi agents-based simulations (MABS) on a cluster or a grid in order to generate sufficient data for scientific use (e.g. in the case of a sensibility analysis of a simulation). We provide a new, general and user-friendly approach to marry MABS and High-Performance Computing (HPC). We, thus, propose a workflow and an associated HPC infrastructure. These two permit to easily deploy a lot of simulations on a cluster without any prior parallelizing work. The method wants to be as generic as possible: no particular MABS targeted, no overhead and HPC compliance work has to be done only once. Moreover the user is guided by a web interface that handles the workflow.

Using purely agent-based platforms for any kind of simulation requires to address the following challenges: (1) scalability (efficient scheduling of agent cycles is difficult), (2) efficient memory management (when and which data should be fetched, cached, or written to / from disk), and (3) modelling (no generally accepted meta-models exist: what are essential concepts, what implementation details?). While dedicated professional simulation tools usually provide rich domain libraries and advanced visualisation techniques, and support the simulation of large scenarios, they do not allow for "agentization" of single components. We are trying to bridge this gap by developing a distributed, scalable runtime platform for multiagent simulation, MASeRaTi, addressing the three problems mentioned above. It allows to plug-in both dedicated simulation tools (for the macro view) as well as the agentization of certain components of the system (to allow a micro view). If no agent-related features are used, its performance should be as close as possible to the legacy system used. Paper type: Technological or Methodological 1 To agentize means to transform given legacy code into an agent so that it belongs to a particular multiagent system (MAS). This term was coined in [29]. In [28], Shoham used the term agentification for this.

2011

Real-life city-traffic simulation presents a good example of multi-agent simulations involving a large number of agents (each human modelled as an individual agent). Analysis of emergent behaviors in social simulations largely depends on the number of agents involved (more than 100,000 agents at least). Due to large number of agents involved, it takes several seconds or even minutes to simulate a single second of real-life. Hence, we resort to distributed computing to speed-up the simulations. We used Hadoop to manage and execute multi-agent application on a small cloud of computers and used CUDA framework to develop an agent-simulation model on GPUs. Hadoop is known for efficiently supporting massive work-loads on a large number of systems with the help of its novel task distribution and the MapReduce model. It provides a fault-tolerant and failure-resilient underlying framework and these properties get propagated to the multi-agent simulation solution developed on top of it. Furth...

Proceedings of the IEEE, 2001

Agent-based systems are increasingly being applied in a wide range of areas including telecommunications, business process modeling, computer games, control of mobile robots, and military simulations. Such systems are typically extremely complex and it is often useful to be able to simulate an agent-based system to learn more about its behavior or investigate the implications of alternative architectures. In this paper, we discuss the application of distributed discrete-event simulation techniques to the simulation of multiagent systems. We identify the efficient distribution of the agents' environment as a key problem in the simulation of agent-based systems and present an approach to the decomposition of the environment that facilitates load balancing.

2013 Winter Simulations Conference (WSC), 2013

This paper describes a holistic architecture for super real-time multi-agent simulation platforms by implementing a complete and integrated simulation stack including a simulation runtime and an application layer that can be used for such situations as traffic simulations. With our prototype system, the first experiment tested the performance scalability when simulating millions of agents on 1,536 CPU cores over 256 nodes. By compiling the X10-based agent simulation system into C++ using MPI, we could run 600 simulation steps in only 78 seconds, which is nearly 10 times faster than real time. We then tested a national network spanning Japan, which was able to simulate a 100 million agents and at near-real time while using 128 nodes. This was the first attempt to deal with such a large number of agents and shows that this infrastructure could be used for large-scale agent simulations in various fields.

Lecture Notes in Computer Science, 2004

Large scale agent-based simulation has become the focus of great interest from both academy and industry in resent years. It has been shown an effective tool for understanding a variety of complex systems, such as market economics, war games, and epidemic propagation models. As the systems of interest grow in complexity, it is often desirable to have different categories of artificial agents that execute tasks on different time scales. With the added complexity, the scalability of a simulation environment becomes a crucial measure of its ability in coping with the complexity of the underlying system. In this paper, we present the design of Samas, a highly scalable architecture for multi-resolution agent-based simulation. Samas is a dynamic data driven application system(DDDAS) that adopts an efficient architecture that allows large number of agents to operate on a wide range of time scales. It uses gossiping, an efficient broadcasting communication model for maintaining the overall consistency of the simulation environment. We demonstrate the effectiveness of this communication model using experimental results obtained from simulation.

Proceedings of the 2006 Winter Simulation Conference, 2006

The PDES literature offers a rich set of techniques for distributed and efficient simulation. However, there is a growing need for simulators that support agent-based applications, and PDES systems are not always well suited for these applications. Example agent-based applications include simulation of biological systems such as ants and bees, multi-robot systems and battlefield simulations. The robotics research community has developed agent-based simulators that provide useful APIs for agent applications. However, such simulators have performance limitations, and they do not scale well. Our approach is to provide middleware between an agent-based API and a PDES simulation kernel. The result is a simulation system that offers an agentbased API for the programmer to a high performance PDES system. Here we describe our design and initial implementation of SASSY, the Scalable Agents Simulation System. We describe our initial implementation and compare the design with related approaches.

Lecture Notes in Computer Science, 2014

Large scale distributed agent-based simulations run on several computing units (e.g., virtual machines in the Cloud, nodes in a supercomputer). Classically, these systems try to (re-)load-balance the nodes as overloaded nodes slow down the process.However another challenge in large scale distributed simulations is that the overall load evolves. In this paper we leverage on commodity computing to adapt resource provisioning (number of computing units) to the load during the execution of the simulation. We also propose an asynchronous migration mechanism that migrate workload between computing nodes efficiently when nodes wait for synchronisation barriers to happen. We validate our implementation on a scenario simulating one day of vehicular traffic in Tokyo, running on 2 to 8 machines depending on the demand. Our evaluation shows a 26% reduction in data migration time compared to a naive migration approach between computing units.

International Journal of High Performance Computing and Networking, 2009

The application of distributed simulation techniques to ever more complex problems calls for more computational power. The emergence of the Grid provides an opportunity to address this challenge and thus make advances in the modelling and analysis of large-scale systems by harnessing the power of many computers. These requirements are met by the design and development of frameworks that support the integration of appropriate Distributed Simulation Kernels (DSK) and Grid middleware. This paper presents a framework and a prototype implementation for specifying and executing distributed simulations over the Grid. The framework incorporates a well established DSK with the Globus middleware.