Probabilistic grid scheduling
based on job statistics and monitoring information.
Doctoral thesis, University of London.
This transfer thesis presents a novel, probabilistic approach to scheduling applications on computational Grids based on their historical behaviour, current state of the Grid and predictions of the future execution times and resource utilisation of such applications. The work lays a foundation for enabling a more intuitive, user-friendly and effective scheduling technique termed deadline scheduling. Initial work has established motivation and requirements for a more efficient Grid scheduler, able to adaptively handle dynamic nature of the Grid resources and submitted workload. Preliminary scheduler research identified the need for a detailed monitoring of Grid resources on the process level, and for a tool to simulate non-deterministic behaviour and statistical properties of Grid applications. A simulation tool, GridLoader, has been developed to enable modelling of application loads similar to a number of typical Grid applications. GridLoader is able to simulate CPU utilisation, memory allocation and network transfers according to limits set through command line parameters or a configuration file. Its specific strength is in achieving set resource utilisation targets in a probabilistic manner, thus creating a dynamic environment, suitable for testing the scheduler’s adaptability and its prediction algorithm. To enable highly granular monitoring of Grid applications, a monitoring framework based on the Ganglia Toolkit was developed and tested. The suite is able to collect resource usage information of individual Grid applications, integrate it into standard XML based information flow, provide visualisation through a Web portal, and export data into a format suitable for off-line analysis. The thesis also presents initial investigation of the utilisation of University College London Central Computing Cluster facility running Sun Grid Engine middleware. Feasibility of basic prediction concepts based on the historical information and process meta-data have been successfully established and possible scheduling improvements using such predictions identified. The thesis is structured as follows: Section 1 introduces Grid computing and its major concepts; Section 2 presents open research issues and specific focus of the author’s research; Section 3 gives a survey of the related literature, schedulers, monitoring tools and simulation packages; Section 4 presents the platform for author’s work – the Self-Organising Grid Resource management project; Sections 5 and 6 give detailed accounts of the monitoring framework and simulation tool developed; Section 7 presents the initial data analysis while Section 8.4 concludes the thesis with appendices and references.
|Title:||Probabilistic grid scheduling based on job statistics and monitoring information|
|Open access status:||An open access version is available from UCL Discovery|
|Additional information:||This is a transfer thesis for the degree of Doctor of Philosophy.|
|UCL classification:||UCL > School of BEAMS > Faculty of Engineering Science > Electronic and Electrical Engineering|
Archive Staff Only