100 likes | 228 Views
Performance Evaluation of an Agent-based Resource Management Infrastructure for Grid Computing. Junwei Cao Darren J. Kerbyson Graham R. Nudd. Department of Computer Science University of Warwick. Grid Resource Management. Globus. Requirements Heterogeneity
E N D
Performance Evaluation of anAgent-based Resource Management Infrastructure for Grid Computing Junwei Cao Darren J. Kerbyson Graham R. Nudd Department of Computer Science University of Warwick
Grid Resource Management Globus Requirements • Heterogeneity • multiplicity of resources and numerous administrative domains • Scalability • millions of resources across wide geographical distances. • Adaptability • resource failure, performance change, etc Legion Condor Ninf NetSolve
Agent-Based Methodology • Agent • A representation of computing resources in the metacomputing environment. • Both a resource provider and a resource requestor. • A router (or matchmaker) between an application and an available resource. • Organised into a hierarchy. • Resource • The detail information of a resource within the grid. • Request • The detail information of an application from the user.
1 Get 2 Get 1 AppInfo 3 AppInfo R/A R/A A A U/A U/A 3 ResInfo 2 ResInfo 4 Return 4 Return Resource Discovery • Resource Advertisement • The resource information can be advertised in the agent hierarchy (both up and down). • Resource Discovery • The application information from the user can be transferred in the agent hierarchy to discover an available resource. Data-pull • Strategies: • No resource advertisement, then complex resource discovery. • Full resource advertisement, then no resource discovery. Data-push
Performance Metrics • Discovery Speed • System Efficiency • Load balancing • Success Rate
Performance Optimisation Strategies Vary by • Dynamics • Agent structure • Resource distribution • Pre-knowledge Caching resource info Using local resource info Using global resource info Limit scope Limit resource validation
A4 Simulator - Modelling • Input to model • Agent system structure • Request distribution • Resource distribution • Performance optimisation strategies • Modelling level • Agent-level (each individuals) • System-level (global)
A4 Simulator - Simulation • Full support for all performance metrics • Multi-view simulation results • Each step view • Accumulative view • Agent View • Log view • Dynamic simulation result display • Comparing strategies
A Case Study • Choice of strategies >> higher performance • No resource advertisement >> low discovery speed low efficiency • Too much resource advertisement >> extreme high discovery speed extreme low efficiency • Reasonable resource advertisement >> high discovery speed high efficiency
Ongoing Work - ARMS • ARMS: an Agent-based Resource Management System for grid computing • A hierarchy of homogenous agents with resource discovery capabilities as meta-level resource management • PACE (a Performance Analysis and Characterisation Environment) as local resource scheduler.