Parallel Optimisations of Perceived Quality Simulations

Abstract

Processor architectures have changed significantly, with fast single core processors replaced by a diverse range of multicore processors. New architectures require code to be executed in parallel to realize these performance gains. This is straightforward for small applications, where analysis and refactoring is simple or existing tools can parallelise automatically, however the organic growth and large complicated data structures common in mature industrial applications can make parallelisation more difficult.
One such application is studied, a mature Windows C++ application used for the visualisation of Perceived Quality (PQ). PQ simulations enable the visualisation of how manufacturing variations affect the look of the final product. The application is commonly used, however suffers from performance issues. Previous parallelisation attempts have failed.
The issues associated with parallelising a mature industrial application are investigated. A methodology to investigate, analyse and evaluate the methods and tools available is produced. The shortfalls of these methods and tools are identified, and the methods used to overcome them explained. The parallel version of the software is evaluated for performance. Case studies centring on the significant use cases of the application help to understand the impact on the user.
Automated compilers provided no parallelism, while the manual parallelisation using OpenMP required significant refactoring. A number of data dependency issues resulted in some serialised code. Performance scaled with the number of physical cores when applied to certain problems, however the unresolved bottlenecks resulted in mixed results for users. Use in verification did benefit, however those in early design stages did not. Without tools to aid analysis of complex data structures, parallelism could remain out of reach for industrial applications. Methods used here successfully, such as serialisation, and code isolation and serialisation, could be used effectively by such tools.