Optimizing the Performance of Directive-based Programming Model for GPGPUs

Accelerators have been deployed on most major HPC systems. They are considered to improve the performance of many applications. Accelerators such as GPUs have an immense potential in terms of high compute capacity but programming these devices is a challenge. OpenCL, CUDA and other vendor-specific models for accelerator programming definitely offer high performance, but these are low-level models that demand excellent programming skills; moreover, they are time consuming to write and debug. In order to simplify GPU programming, several directive-based programming models have been proposed, including HMPP, PGI accelerator model and OpenACC. OpenACC has now become established as the de facto standard. We evaluate and compare these models involving several scientific applications. To study the implementation challenges and the principles and techniques of directive- based models, we built an open source OpenACC compiler on top of a main stream compiler framework (OpenUH as a branch of Open64). In this dissertation, we present the required techniques to parallelize and optimize the applications ported with OpenACC programming model. We apply both user-level optimizations in the applications and compiler and runtime-driven optimizations. The compiler optimization focuses on the parallelization of reduction operations inside nested parallel loops. To fully utilize all GPU resources, we also extend the OpenACC model to support multiple GPUs in a single node. Our application porting experience also revealed the challenge of choosing good loop schedules. The default loop schedule chosen by the compiler may not produce the best performance, so the user has to manually try different loop schedules to improve the performance. To solve this issue, we developed a locality-aware auto-tuning framework which is based on the proposed memory access cost model to help the compiler choose optimal loop schedules and guide the user to choose appropriate loop schedules.