DS-Grid: Large Scale Distributed Simulation on the Grid Georgios Theodoropoulos Midlands e-Science Centre University of Birmingham, UK Stephen John Turner, Wentong Cai Parallel & Distributed Computing Centre Nanyang Technological University, Singapore Brian Logan University of Nottingham, UK
2/20 e-Science Workshop Outline MeSC and the DS-Grid Project Motivation & Challenges HLA_Grid –Benchmark Experiments HLA_Grid_RePast –Large Scale Agent Based Simulation –Experiments and Results Conclusions and Future Work
3/20 e-Science Workshop MeSC: Centre of Excellence – Modelling and Simulation of Large Complex Systems Funded by the UK e-Science programme Part of the national Grid infrastructure of the UK Virtual centre with a base in the School of Computer Science
4/20 e-Science Workshop The DS-Grid Project One of only four “Sister Projects” funded by the e- Science Core Programme UK-Singapore Grid link
5/20 e-Science Workshop Motivation The development of complex simulation applications usually requires collaborative effort from researchers with different domain knowledge and expertise, possibly at different locations These simulation systems often require huge computing resources and the data sets required by the simulation may also be geographically distributed The Grid offers an unrivalled opportunity : –Enables collaboration –Enables the use of distributed computing resources, –Allows access to geographically distributed data sets –Supports service-oriented architectures that can facilitate model and resource discovery
6/20 e-Science Workshop DS-Grid Vision A Grid “plug-and-play” distributed collaborative simulation environment, where researchers with different domain knowledge and expertise, at different locations, develop, modify, assemble and execute distributed simulation components over the Grid
7/20 e-Science Workshop High Level Architecture Interface Run-Time Infrastructure (RTI) Federation ManagementDeclaration Management Object ManagementOwnership Management Time ManagementData Distribution Management Passive Viewers Simulations Simulation Surrogates SOM FED FOM HLA Rules (Federations) HLA Rules (Federates) Federation
8/20 e-Science Workshop HLA and the Grid Discovery of Models Discovery of Resources Management of Simulation Execution Model Factory federate RTI
9/20 e-Science Workshop Challenges Model Discovery and Matching –While HLA provides interoperability at the communication level there is little support for interoperability at the semantic level Resource Management –HLA does not provide support for resource management and dynamic load balancing Simulation Management on the Grid –HLA does not provide any support for collaborative development of simulation components –New Grid-aware collaborative environments for distributed simulation must be developed
10/20 e-Science Workshop Client Simulation Code Grid-aware HLA API Globus Resource Proxy Grid-aware HLA API Proxies & Federates HLA API RTI on LAN Grid Network Grid Services Grid Services: indexing, discovery, resource management, monitoring services … Globus HLA API HLA_Grid
11/20 e-Science Workshop Experimental Environment
12/20 e-Science Workshop ► Overhead in cluster: latency = 50 millisecond ► Use of GT3, encoding/decoding of parameters/results, and the communication costs ► Overhead in WAN: latency = 1150 millisecond ► Mainly caused by the increase in communication using SOAP messages over long distances -> increase number of packets Benchmark Experiments
13/20 e-Science Workshop HLA_Grid_RePast Executes distributed, large scale simulations of agent-based systems over the Grid Integrates HLA_Grid and RePast (Java based toolkit for lightweight agents) Each federate divided into two parts: –Client Side RePast Code and HLA_Grid Library –Remote Side Proxy RTI Ambassador and Federate Proxy Ambassador
14/20 e-Science Workshop Structure of HLA_Grid_Repast
15/20 e-Science Workshop Case Study: Tileworld
16/20 e-Science Workshop Network Configuration for Experiments
17/20 e-Science Workshop Experimental Results Performance on a LAN (PC Cluster) with one agent federate
18/20 e-Science Workshop Experimental Results Performance on a WAN (Grid) with one agent federate
19/20 e-Science Workshop Conclusions Advantages –Avoids many firewall issues as client communicates with proxy via Grid services –Enables easier integration with non HLA simulators –Hierarchical federations may be constructed easily –Provides easy migration of client code as proxy does not need to be migrated Disadvantages –Overhead of communication as all simulation events use Grid services
20/20 e-Science Workshop Future Work Further analysis of communication network traffic Additional Case Studies Grid “plug-and-play” Environment based on Service- Oriented Architecture –Component Based Simulation Development Service Discovery Semantic Matching –Efficient Execution Resource Management –Collaborative Environment for Distributed Simulation Workflow Management