|
 |
||||||||
| Department of Computer Science, University of Edinburgh |
ECoVism is a three project funded by EPSRC, in collaboration with the University of Manchester. The technical aims of the project are: to provide a unified framework for memory consistency; to develop appropriate compiler analysis that can utilise this framework effectively; and, thereby, to demonstrate automatic compilation of optimised applications for virtual shared memory (VSM) systems. The measurable goals are: to approach the efficiency of hand-crafted message-passing on state-of-the-art non-uniform memory access (NUMA) architectures; and to investigate the performance impact this will have on future generation hardware architectures.
SummaryParallel computers of increasing size are becoming ever more widely available as a potential means for achieving high performance-critical applications. However, there is continuing debate about the best form of implementation (hardware/software) for this purpose; in particular, the relative merits and problems of writing programs using the methods of 'data sharing' (in a single-address-space) versus 'message-passing' (between multiple-address-spaces) are strongly disputed. This project will design, implement and demonstrate novel virtual shared memory (VSM) schemes which aim to achieve the advantages of both the above methods while eliminating their disadvantages. The work will firstly develop suitable formal methods of (hardware/firmwae) schemes for memory consistency. These models will then be used as a basis for design and implementation of compiler analysis and consequent selection (during parallel code generation) of the most appropriate options for a specific platform. The efficacy of the resulting system will be assessed by means of extensive testing on a sample of performance-critical applications across a range of simulated and actual platforms. The research will be performed in collaboration with three groups in, or closely associated within, leading compter system vendors.