Application acceleration : an investigation of automatic porting methods for application accelerators

Future HPC systems will contain both large collections of multi-core proces sors and specialist many-core co-processors. These specialised many-core co processors are typically classified as Application Accelerators. More specifically, Application Accelerations are devices such as GPUs, CELL Processors, FPGAs and custom application specific integrated circuit devices(ASICs). These devices present new challenges to overcome, including their programming difficulties, their diversity and lack of commonality of programming approach between them and the issue of selecting the most appropriate device for an application. This thesis attempts to tackle these problems by examining the suitability of automatic porting methods. In the course of this research, relevant software, both academic and com mercial, has been analysed to determine how it attempts to solve the problems relating to the use of application acceleration devices. A new approach is then constructed, this approach is an Automatic Self-Modifying Application Porting system that is able to not only port code to an acceleration device, but, using performance data, predict the appropriate device for the code being ported. Additionally, this system is also able to use the performance data that are gathered by the system to modify its own decision making model and improve its future predictions. Once the system has been developed, a series of applications are trialled and their performance, both in terms of execution time and the accuracy of the systems predictions, are analysed. This analysis has shown that, although the system is not able to flawlessly predict the correct device for an unseen application, it is able to achieve an accuracy of over 80% and, just as importantly, the code it produces is within 15% of that produced by an experienced human programmer. This analysis has also shown that while automatically ported code performs favourably in nearly all cases when compared to a single-core CPU, automatically ported code only out performs a quad-core CPU in three out of seven application case studies. From these results, it is also shown that the system is able to utilise this performance data and build a decision model allowing the users to determine if an automatically ported version of their application will provide performance improvement compared to both CPU types considered. The availability of such a system may prove valuable in allowing a diverse range of users to utilise the performance supplied by many-core devices within next generation HPC systems