30.07.2015 Views

Actas JP2011 - Universidad de La Laguna

Actas JP2011 - Universidad de La Laguna

Actas JP2011 - Universidad de La Laguna

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

<strong>Actas</strong> XXII Jornadas <strong>de</strong> Paralelismo (<strong>JP2011</strong>) , <strong>La</strong> <strong>La</strong>guna, Tenerife, 7-9 septiembre 2011Achieving interactive multiagent simulationsover Jason through Java tuningVíctor Fernán<strong>de</strong>z-Bauset, Francisco Grimaldo, Miguel Lozano y Juan M. Orduña 1Resumen— Java-based simulation environments arecurrently used by many multiagent systems (MAS),since they mainly provi<strong>de</strong> portability as well as aninteresting reduction of the <strong>de</strong>velopment cost. However,this kind of MAS are rarely consi<strong>de</strong>red when <strong>de</strong>velopinginteractive applications with time responseconstraints. This paper analyses the performance provi<strong>de</strong>dby Jason, a well-known Java-based MAS platform,as a suitable framework for <strong>de</strong>veloping interactivemultiagent simulations. We show how to tuneboth the heap size and the garbage collection of theJava Virtual Machine in or<strong>de</strong>r to achieve a good performancewhile executing a simple locomotion benchmarkbased on crowd simulations. Furthermore, thepaper inclu<strong>de</strong>s an evaluation of Jason’s performanceover multi-core processors. The main conclusion <strong>de</strong>rivedfrom this work is that, by means of Java tuning,it is possible to run interactive MAS programmed usingJason.Palabras clave— Jason, Java, simulation, interactive,multiagent.I. Introduction and Related workMAS platforms capable of handling a largeamount of complex autonomous agents at interactiveresponse times are required by interactivemultiagent applications such as crowd simulationsand massive online games. Usually, these kinds ofsimulations involve a high number of agents (e.g.pe<strong>de</strong>strians) interacting in a shared environment. Interactivity,in turn, requires the use of parallel techniquesthat allow to validate and to execute the actionsrequested within a limited period of time (commonly,250 ms [1]).Java-based simulation environments are currentlybeing used by many MAS, since they mainly provi<strong>de</strong>portability as well as an interesting reductionof the <strong>de</strong>velopment cost. However, this kind of MASare rarely consi<strong>de</strong>red when <strong>de</strong>veloping interactive applicationswith time response constraints, becauseof Java being normally less efficient than other languagessuch as C or C++. This situation requestsperforming a specific Java tuning to be able to tacklethis type of applications. In this paper, we show theJava tuning carried out for the purpose of evaluatingthe performace of Jason [2], a well-known Java-basedMAS platform. The aim of this tuning is to adjustboth the heap size and the garbage collection of theJava Virtual Machine in or<strong>de</strong>r to satisfy the temporalrequirements of interactive multiagent simulations.Therefore, the results presented in this paper willalso be of great value to those researches consi<strong>de</strong>r-1 Computer Science Department, University ofValencia, Dr. Moliner 50, (Burjassot) Valencia,Spain. E-mail: victor.fernan<strong>de</strong>z-bauset@uv.es,francisco.grimaldo@uv.es, miguel.lozano@uv.es,juan.orduna@uv.es.ing Java-based simulation environments suitable for<strong>de</strong>veloping interactive multiagent applications.When <strong>de</strong>veloping this kind of interactive MASthree layers are normally consi<strong>de</strong>red: the computerarchitecture, the MAS platform and the graphical engine(if any). At the low level, different distributedcomputer architectures have been applied in or<strong>de</strong>r toallow massive interactive simulations to scale up withthe number of agents by simply adding new hardware(e.g. networked-server, P2P, etc.). For instance,a new approach has been presented for PLAYSTA-TION3 which supports simulation of simple crowdsof up to 15000 individuals at 60 frames per second[3]. Parallel simulation, based on classical Reynolds’sboids [4], has been also integrated in a PC-Clusterwith MPI communication [5] to finally produce smallsimulations (512 boids). At the top level, the graphicalengine of the application must ren<strong>de</strong>r the visualizationat interactive frame rates. The computergraphics community generally represents the MASas a particle system with local interactions [6], [7],though, few works inclu<strong>de</strong> socially complex and autonomousbehaviors [8]. However, they are not normallybased on standard agent architectures.In the middle level, the MAS platform is in chargeof providing the required data flow to the graphicalengine while efficiently using the computationalresources. Thus, it constitutes a key middlewarethat highly influences the global performance andthe scalability of the system. It mainly addressestwo important issues: mo<strong>de</strong>ling the behavior of theagents as well as their parallel lifecycle execution.Java is a popular language providing built-in supportfor concurrency that is commonly used by MAS platforms.Although Java performance has been studiedfrom different perspectives, probably the mostusual is to tune server applications running on largemulti-processor servers [9]. There are more specificworks focused on the evaluation of Java-based multiagentplatforms [10], [11], [12]. However, none ofthem <strong>de</strong>als with providing interactivity to the correspondingMAS. Some researchers have been also testingthe performance and scalability of a few existingMAS platforms [13], showing a lack of both importantissues in many of them. In a previous work [14],the authors analysed Jason’s architecture and evaluatedits performance un<strong>de</strong>r both centralised and distributedinfrastructures. Regardless the infrastructure,the results showed that the execution optionshad to be reviewed in or<strong>de</strong>r to achieve a more equilibratedresponse time distribution, an aspect that wehave covered in this work.The rest of the paper is organized as follows. Sec-<strong>JP2011</strong>-607

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!