Author(s): Parastatidis S
Abstract: Advance in hardware technologies combined with decreased costs have started a trend towards massively parallel architectures that utilise commodity components. It is though unreasonable to expect software developers to manage the high degree of parallelism that is made available by these architectures. This thesis argues that a new programming model is essential for the development of parallel applications and presents a model which embraces the notions of object-orientation and implicit identification of parallelism. The new model allows software engineers to concentrate on development issues, using the object-oriented paradigm, whilst being freed from the burden of explicitly managing parallel activity. To support the programming model, the semantics of an execution model are defined and implemented as part of a run-time support system for object-oriented parallel applications, Details of the novel techniques from the run-time system, in the areas of lazy task creation and object-based, distributed shared memory, are presented. The tasklet construct for representing potentially parallel computation is introduced and further developed by this thesis. Three caching techniques that take advantage of memory access patterns exhibited in object-oriented application are explored. Finally, the performance characteristics of the introduced run-time techniques are analysed through a number of benchmark applications.
Notes: British Lending Library DSC stock location number: DXN038005