high performance computing on graphics processing units: hgpu.org

Scientific applications have an increasing need of resources and many grand scientific challenges require exascale compute capabilities to be addressed. One major concern to achieve exascale is programmability. New automatic methods are required to fill the gap between developers of scientific applications and HPC experts. In addition, as scientific applications are becoming more and more complex and are supposed to run at extreme scale, new tools are required to assist developers in the debugging phase of application development. This thesis explores the combination of static and dynamic methods to improve programmability of HPC applications. Two major issues are investigated: the complexity of programming heterogeneous architectures and the prevention of deadlocks in parallel programs. The first part of this thesis investigates the automatic task adaptation for heterogeneous architectures. More precisely, we propose a new method to improve programmability of heterogeneous architectures. The programmer expresses the parallelism of his application through a sequence of OpenCL tasks without considering issues related to the underlying architecture where its code will be executed. Then our method automatically partitions the tasks into sub-tasks executed by each device and handles load balancing between the devices to take full advantage of the machine capabilities. The second part of this thesis investigates the automatic detection and prevention of deadlocks in parallel programs. We propose a novel static analysis to precisely detect execution paths in parallel programs potentially leading to deadlocks. This static analysis is then combined with a dynamic instrumentation of the code to automatically prevent deadlocks at runtime. The solutions proposed in this thesis have been tested and validated on real parallel applications.