University Homepage School Homepage School Contacts School Search

PhD Scholarship: Very Large Scale Configuration for Grid Computing

Applications are invited from UK students for a three year PPARC eScience Studentship, commencing in October 2002.  Details can be found:

http://www.pparc.ac.uk/rs/fs/es/esciencestudentships02.asp

The student will be based in the Division of Informatics at Edinburgh University, and will involve an investigation of the problems associated with specification and implementation of configurations for large scale computing fabrics.

Project description:

No existing configuration system is proven at the scale of 10,000 or more computing nodes.  No commercial systems address these issues in a scalable manner and academic research has no solution that is close to being tested.  Yet combined server farms of this size are already being planned for the next few years within the emergent field of grid computing.  There is a growing consensus that solving configuration problems on this scale will require different techniques to those hitherto deployed.

The effective management of fabrics of this size and larger is essential to support most large-scale experimental efforts over the next 10-20 years.  Solving the fabric management problem is an essential pre-requisite to most PPARC-funded science (e.g., GridPP, EU DataGrid and AstroGrid will all have acute fabric management problems unless these issues are dealt with).  Without a scalable approach to management problems we do not see a way to build and maintain grid computing facilities in a cost-effective way.

The special challenges of large scale configuration include

• High level description languages.  Procedural representations of configuration data are unworkable at this scale.  Declarative languages with support for high level abstractions, automated verification and consistency checks, and explicit semantics, are largely lacking at present.
• Dynamics and resiliency.  Tight central control of every property of every node becomes infeasible as scale increases.  Nodes need be able to perform some degree of self-configuration autonomously, thereby introducing `dynamic' configuration parameters -- and increasing resiliency.  How can this be integrated with centrally held, static, configuration representations?
• Synchronisation and sequencing.  One aspect of central control is the temporal ordering of configuration changes.  Novel approaches are required to make this manageable in the face of massive scale.
• Configuration changes as transactions.  Commit-rollback transactions, versioning and access control are desirable properties of large scale configuration systems.
• Modelling intrusiveness.  A computing fabric composed of thousands of nodes may be running thousands of jobs for hundreds of users.  Making configuration changes causes varying degrees of disturbance to these jobs and their owners.  But making no changes is not an option.  What is the right model for `intrusiveness'?

These are relatively well defined problems. And partial solutions to some of them already exist -- recent work, for example, has shown how "asymptotic configuration" can be an effective technique for very large numbers of nodes.  The PhD candidate will experiment with solutions to large-scale configuration problems on the testbed systems already under development in Edinburgh to investigate the problems and map out the solution space for large-scale configuration system.  This work will be coupled to more principled work on the design of configuration description systems based on new models of configuration management for large-scale systems.

The candidate would be able to build on seven years of experience in the Division of Informatics: both with designing and implementing a successful medium scale configuration system (LCFG), and with relying on it in a demanding production environment.

The student will also work closely with colleagues in EPCC and the National e-Science Centre to connect to the large applications funded by PPARC (most notably GridPP and AstroGrid).  Close collaboration with these projects and the EU DataGrid project will provide strong application drivers to shape and inform the work.

Collaborations:

Informatics at Edinburgh is involved in day-to-day collaboration with CERN, as the Edinburgh LCFG system has been selected as the underlying configuration technology for the European DataGrid project.  We also have fruitful links with HP Labs at Bristol, and other informal contacts in the large scale configuration domain.

Informatics and EPCC at Edinburgh are core members of the consortium running the National e-Science Centre.  Our work in the Centre provides very good links to ScotGrid (a prototype EU DataGrid tier 2 centre); AstroGrid (co-ordinated by colleagues at RoE); and GridPP co-ordinated by our colleagues in Glasgow department of Physics).  This proposal both exploits PPARC funding of configuration work at Edinburgh and adds value to the existing PPARC funded RA working on configuration management in Edinburgh.

Applications can be obtained from

Betty Hughes
Admin Secretary
Graduate School
Division of Informatics
University of Edinburgh
2 Buccleuch Place
Edinburgh EH8 9LW

or by e-mail phd-admissions@informatics.ed.ac.uk, specifically mentioning the above studentship in the comments field. All completed applications should be received by the Graduate School (address above) by 31st May 2002.  Any informal enquiries may be made to Paul Anderson paul@dcs.ed.ac.uk.