A REST Model for High Throughput Scheduling in Computational Grids

Current grid computing architectures have been based on cluster management and batch queuing systems, extended to a distributed, federated domain. These have shown shortcomings in terms of scalability, stability, and modularity. To address these problems, this dissertation applies architectural styles from the Internet and Web to the domain of generic computational grids. Using the REST style, a flexible model for grid resource interaction is developed which removes the need for any centralised services or specific protocols, thereby allowing a range of implementations and layering of further functionality. The context for resource interaction is a generalisation and formalisation of the Condor ClassAd match-making mechanism. This set theoretic model is described in depth, including the advantages and features which it realises. This RESTful style is also motivated by operational experience with existing grid infrastructures, and the design, operation, and performance of a proto-RESTful grid middleware package named DIRAC. This package was designed to provide for the LHCb particle physics experiment's â€ﾜoff-line” computational infrastructure, and was first exercised during a 6 month data challenge which utilised over 670 years of CPU time and produced 98 TB of data through 300,000 tasks executed at computing centres around the world. The design of DIRAC and performance measures from the data challenge are reported. The main contribution of this work is the development of a REST model for grid resource interaction. In particular, it allows resource templating for scheduling queues which provide a novel distributed and scalable approach to resource scheduling on the grid.