|
| EPSRC Reference: |
GR/N25220/01 |
| Title: |
ABSTRACTION & PERFORMANCE: DESIGN AND IMPLEMENTATION OF CATEGORICAL LANGUAGES FOR HIGH PERFORMANCE COMPUTIN |
| Principal Investigator: |
Professor AD Kennedy |
| Other Investigators: |
|
| Researcher Co-investigator: |
|
| Project Partner: |
| E P C C Ltd |
Numerical Algorithms Group Ltd |
|
|
| Department: |
Sch of Physics and Astronomy |
| Organisation: |
University of Edinburgh |
| Scheme: |
Standard Research |
| Starts: |
01 September 2000 |
Ends: |
31 August 2003 |
Value (£): |
62,975
|
| EPSRC Research Topic Classifications: |
|
| EPSRC Industrial Sector Classifications: |
|
| Related Grants: |
|
| Panel History: |
|
|
Summary |
The aim of this project is to investigate the design and implementation of categorical languages allowing high levels of abstraction, portability and efficient implementation. It is to be based on the existing language Aldor and its implementations: and it will address in particular the implementation of the algorithms used for computations of Quantum Chromodynamics (QCD) on lattice. This project will form a feasibility study into whether it is possible to extract sufficient performance without losing abstraction for scientific computing.
|
| Final Report Summary |
The aim of this project is to investigate the design and implementation of categorical languages allowing high levels of abstraction, portability and efficient implementation. It is to be based on the existing language Aldor and its implementations: and it will address in particular the implementation of the algorithms used for computations of Quantum Chromodynamics (QCD) on lattice. This project will form a feasibility study into whether it is possible to extract sufficient performance without losing abstraction for scientific computing.
NEW ABSTRACT
T.Ashby carried out the main part of the project, and his work will be written up in his Ph.D thesis (in preperation). He presented his preliminary results at the Third International Workshop on Numerical Analsysis and Lattice QCD held at the National e-Science Centre, 30th June - 4th July 2003. In addition, he presented some results at the Aldor Users' Group Meeting at UWO in Ontario, Canada on 26th July 2001. His work concentrated on formulating Krylov space techniques for solving large sparse linear systems in a modular and unified way, and showing that the potential inefficiencies introduced by this modular approach could be avoided by carefull use of lazy evaluation together with copious code inlining. He has carried out detailed numerical experiments to test how well this approach works. The problem is highly non-tricial, as it not only lies at the intersection of state-of-the-art research in computer science and theoretical Physics, but also in using sophisticated state-of-the-art Numerical Analysis techniques. The fact that it was presented at a meeting of QCD experts and world-leading Numerical Analysts and was received with great interst by both may be taken as an indication of its quality.
An auxilliary part of the project was the development of a prototype Aldor code for some key kernels for QCD computations, carried by T.AShby, D.Galletly, B, Joe, A.D. Kenned, and G Lacagnina. This demonstrated that it is indeed possible to write modular portable and efficient codes as proposed, although the high-performance aspects have not yet been well excercised. In part, this is because, during the course of the project, NAG decided not to support Aldor as a product, and its suport then passed to its original author (now at the University of Western Ontario, Canada). This led the project collaborators to split their efforts in two directions, firstly using Aldor, as discussed in the previous paragraph. The former direction was dictated by the large-scale support for developing such a portable, modular, and efficient QCD code by the SciDAC program of the US DOE.
We believe that the project has not only been successful in demonstrating that largescale numerical codes targeted at massively parallel computers can be written in a modular, portable, yet efficient way using categorical languages; but also has demonstrated a harmonoius and succeful collaboration between theoretical physicists, computer scientists, and numerical analysts. We intend to continue tis line of research, and to this end several proposals are being prepared or have been submitted; e.g. a large European IP project is currently being reviewed by the European Commision.
|
| Further Information: |
|
| Organisation Website: |
http://www.ed.ac.uk |
|
|