Motivation Architecture Algorithms Examples Dynamics Diffusion scheduling in

Motivation Architecture Algorithms Examples Dynamics Diffusion scheduling in multiagent computing system Robert Schaefer, AGH University of Science and Technology, Kraków, Poland The Group Members: Maciej Smołka Jagiellonian University, Kraków, Poland Piotr Uhruski, Marek Grochowski AGH University of Science and Technology, Kraków, Poland

Motivation Architecture Algorithms Examples Dynamics Distributed. Motivation computation paradigms q message passing libraries PVM Parallel Virtual Machine (1990), MPI Message-Passing Interface (1992) q SOA (Service Oriented Architecture) CORBA (1996), SOAP (1998) q GRID Condor (1997), Globus (1998), OGSI/OGSA (2002) Some drawbacks : q partially manual resources allocation q time consuming deployment and maintenance of the system q usually assuming static resources Diffusion scheduling in multiagent computing system

Motivation Architecture Algorithms Examples Distributed computing using Motivation MAS technology Application Computation + Agent logic Middleware Agents environment Heterogeneous Operating Systems Network Diffusion scheduling in multiagent computing system Dynamics

Motivation Architecture Algorithms Examples Dynamics Overview of the OCTOPUS architecture Application Sample task implementations Smart Solid Agents (scheduling, grain control) Middleware VCN Connections Octopus CORBA Java Octopus . . . Virtual Topology Diffusion scheduling in multiagent computing system CORBA Java Agent SDK

Motivation Architecture Algorithms Examples Dynamics OCTOPUS Key Tasks Architecture § Execute Agents § Distributed Communication § Environment Information § Migration § Virtual Network Topology § Virtual Computation Node (VCN) § Agent’s Construction Kit Diffusion scheduling in multiagent computing system Agents environment

Motivation Architecture Algorithms Examples Dynamics Algorithms Diffusion scheduling idea q Analogy to molecular diffusion phenomena q Local scheduling method – every agent is autonomously searching and allocating resources at neighbouring node q We hope to obtain the asymptotically balanced load Diffusion scheduling in multiagent computing system

Motivation Architecture Algorithms Examples Dynamics Diffusion scheduling – main parameters Diffusion schduling in multiagent computing system

Motivation Architecture Algorithms Examples Diffusion scheduling algorithm Diffusion scheduling in multiagent computing system Dynamics

Motivation Architecture Algorithms Examples Dynamics Binding energy formulas under consideration (1) (2) Diffusion scheduling in multiagent computing system

Motivation Architecture Algorithms Examples Dynamics Algorithms Controlling the computation grain – Container agent q Internal job is a dynamic structure of atomic jobs q Sequential computation of contained atomic jobs q New agent creation when the number of contained jobs exceeds the capacity of the agent Diffusion scheduling in multiagent computing system

Motivation Architecture Algorithms Examples Dynamics „Weak” synchronization Algorithms strategy – „Leo the Professional” agent (J. Momot, K. Kossacki – 2004) q Migrates through the network and gathers information about computing agents q Responsible for removing redundancy q Allows to avoid total synchronization of the local system Diffusion scheduling in multiagent computing system

Motivation Architecture Algorithms Examples Dynamics Speedup vs. grain in CAE computation Tests Diffusion scheduling in multiagent computing system

Motivation Architecture Algorithms Examples Dynamics Overhead of the Agent Oriented technology Tests (the case of HGS computation) Diffusion scheduling in multiagent computing system

Motivation Architecture Algorithms Examples Speedup of the Diffusion Scheduling Tests (the case of HGS computation) Diffusion scheduling in multiagent computing system Dynamics

Motivation Architecture Algorithms Examples Dynamics Communication dependent rules „LAN” case Diffusion scheduling in multiagent computing system „WAN” emulation

Motivation Architecture Algorithms Examples Experiments in the local area network (1) Diffusion scheduling in multiagent computing system (2) Dynamics

Motivation Architecture Algorithms Examples Experiments in the wide area network (1) Diffusion scheduling in multiagent computing system (2) Dynamics

Motivation Architecture Algorithms Preliminaries Examples Conclusions Diffusion scheduling in multiagent computing system Dynamics

Motivation Architecture Algorithms Examples State equations Conclusions Diffusion scheduling in multiagent computing system Dynamics

Motivation Architecture Algorithms Examples Optimal scheduling problem Conclusions Diffusion scheduling in multiagent computing system Dynamics

Conclusions • Diffusion scheduling is an effective tool of managing large -scale distributed systems. It is achieved by the low complexity of local scheduling rules and only local communication. It ensures proper agent location in the dynamic network environment. • Introduced formal description provides the discrete equation of evolution and the characterization of admissible controls as well as the cost functional for computing MAS. • The optimal scheduling problem posses the unique solution in the class of stationary strategies. • Total overhead is low in comparison with the computation time (~ 5%). • No significant requirements imposed over applications. Diffusion scheduling in multiagent computing system

Thank you for your patience! Diffusion scheduling in multiagent computing system

Publications