Simulation in the Cloud

Tamas Mahr and Rajmund Bocsi and Laszlo Gulyas

Lorand Eotvos University, Budapest

TITLE: Simulation in the Cloud

ABSTRACT: One of the recent trends on the Internet is the appearance of cloud-based software services. From simple mail clients to entire ERP systems, a whole rage of software is available as a service. In a software as a service setting, being a cloud-based service, the software runs on a remote server (in a cloud), leaving the user's resources free. This allows the user to run software that use more computational resources than available on her local computer. This is typically useful for researchers running agent-based simulation experiments. Due to the bottom-up design of agent-based models, the modeller is destined to explore the behavior of the model during its creation. Emergent properties cannot be deduced from the model, they can only be observed during simulations. The process of exploring a model can be guided by different experiment designs. Nevertheless, it still requires the simulation of the model with multiple diverent parameter settings. The typically high number of parameters and the high number of values of the parameters can make such model exploration (a.k.a. parameter sweep) a resource intensive task. In this talk we present the Model Exploration Service that allows the user to run parameter-sweep experiments in a distributed fashion on a dynamic cluster of computers provisioned on demand. The user is assisted in designing, setting up, and submitting the experiment by a software tool running locally on her computer. The subscription-based simulation service takes care of distributing the experiment on multiple machines in the cloud, and of collecting the results, which can be downloaded by the local tool for further processing. We describe the tool-set that helps the user in accessing the Model Explo- ration Service. This tool-set consists of the extended MEME tool and a web portal. The MEME tool assists the user in designing, setting up, and submitting the experiment in a few clicks. It can handle NetLogo, RepastJ, or FABLES models, or alternatively any Java-based simulation can be extended to enable MEME to handle it. The web portal helps the user to subscribe to the service, and to monitor the running experiments. Possible further improvements of the simulation service include continuous experiment monitoring from the MEME client (as opposed to the portal), en- hanced simulation-time prediction, improved control over running experiments, such as adding or removing compute hosts to an experiment run-time, or defining a cost limit on a per host or per experiment basis.

This work was partially supported by the Hungarian Government, under the grant KMOP-1.1.2-08/1-2008-0002.

Back to Swarmfest2010_program