Strategies and tools for the exploitation of massively parallel computer systems
Evans, Emyr Wyn (2000) Strategies and tools for the exploitation of massively parallel computer systems. PhD thesis, University of Greenwich.
Emyr_Wyn_Evans_2000.pdf - Published Version
Restricted to Repository staff only until 16 March 2017.
Available under License Creative Commons Attribution Non-commercial No Derivatives.
The aim of this thesis is to develop software and strategies for the exploitation of parallel computer hardware, in particular distributed memory systems, and embedding these strategies within a parallelisation tool to allow the automatic generation of these strategies.
The parallelisation of four structured mesh codes using the Computer Aided Parallelisation Tools provided a good initial parallelisation of the codes. However, investigation revealed that simple optimisation of the communications within these codes provided an even better improvement in performance. The dominant factor within the communications was the data transfer time with communication start-up latencies also significant. This was significant throughout the codes but especially in sections of pipelined code where there were large amounts of communication present.
This thesis describes the development and testing of the methods used to increase the performance of these communications by overlapping them with unrelated calculation. This method of overlapping the communications was applied to the exchange of data communications as well as the pipelined communications.
The successful application by hand provided the motivation for these methods to be incorporated and automatically generated within the Computer Aided Parallelisation Tools. These methods were integrated within these tools as an additional stage of the parallelisation. This required a generic algorithm that made use of many of the symbolic algebra tests and symbolic variable manipulation routines within the tools.
The automatic generation of overlapped communications was applied to the four codes previously parallelised as well as a further three codes, one of which was a real world Computational Fluid Dynamics code.
The methods to apply automatic generation of overlapped communications to unstructured mesh codes were also discussed. These methods are similar to those applied to the structured mesh codes and their automation is viewed to be of a similar fashion.
|Item Type:||Thesis (PhD)|
|Uncontrolled Keywords:||massively parallel, computing, hardware, computer software, programming|
|Subjects:||Q Science > QA Mathematics > QA76 Computer software|
|School / Department / Research Groups:||School of Computing & Mathematical Sciences
Faculty of Architecture, Computing & Humanities > School of Computing & Mathematical Sciences
School of Computing & Mathematical Sciences > Parallel Processing Research Group (2000)
Faculty of Architecture, Computing & Humanities > School of Computing & Mathematical Sciences > Parallel Processing Research Group (2000)
|Last Modified:||16 Mar 2016 14:53|
Actions (login required)
Downloads per month over past year