Recent Advances in SmartGridSolve

Oleg GirkoSchool of Computer Science & Informatics

University College Dublin

GridRPC and collective mapping GridRPC limitations

Individual mapping Client-server communication only No communication parallelism

Collective mapping Improved balancing of computation load Reduced volume of communication Improved balancing of communication load Increased parallelism of communication Reduced client memory usage and paging

Collective mapping requirements DAG of task dependencies Order of task execution

SmartGridRPC: runtime discovery

• grpc_map() { … }• Discovery phase and execution phase• Constraints on the code

– No control flow dependent on remote task result

– grpc_local() { … } for side effects

Runtime discovery: fault tolerance

• Error phase• Re-mapping• Re-execution from the beginning• Additional constraints on code

ADL: algorithm definition language

• Separate algorithm definition• Parameterised grpc_map() { … }• Increased programming efforts

– Need to write a separate algorithm definition

– Need to keep ADL in sync with the algorithm– No check if ADL diverged from the algorithm

Static code analysis

• Non-intrusive• Algorithm definition extracted by

code analysis• No limitations of other approaches

– No restrictions on code– No separate algorithm description

Fault tolerance for better performance

• Not restarting from scratch• Keeping log of GridRPC calls and task

dependencies• Restarting failed task and all task it

depends on• Reducing likelihood of data loss

Publications• T. Brady, J. Dongarra, M. Guidolin, A. Lastovetsky, K. Seymour, “SmartGridRPC: The

new RPC model for high performance Grid computing”, Concurrency and Computation: Practice and Experience, vol. 22, issue 18, pp. 2467-2487, 2010

• M. Guidolin, “Performance of GridRPC-based programming systems for distributed scientific computing: issues and solutions”, School of Computer Science and Informatics, Dublin, Ireland, University College Dublin, pp. 168, 04/2010

• M. Guidolin, T. Brady, A. Lastovetsky, “How Algorithm Definition Language (ADL) Improves the Performance of SmartGridSolve Applications”, The 7th High-Performance Grid Computing Workshop, Atlanta, USA, Apr 19, 2010

• O. Girko, A. Lastovetsky, “Using Static Code Analysis to Improve Performance of GridRPC Applications”, 9th High-Performance Grid and Cloud Computing Workshop (HPGC 2012), Shanghai, China, IEEE Computer Society, May 21, 2012

• T. Brady, O. Girko, A. Lastovetsky, “Smart RPC In Grids And Clouds” in “Large Scale Network-Centric Computing Systems”, John Wiley & Sons, to be published in 2012

Questions?