high performance computing on graphics processing units: hgpu.org

The ever increasing complexity of scientific applications has led to utilization of new HPC paradigms such as Graphical Processing Units (GPUs). However, modifying applications to run on GPU is challenging. Furthermore, the speedup achieved by using GPUs has added a huge heterogeneity to HPC clusters. In this dissertation, we enabled NPAIRS, a neuro-imaging application, to run on GPUs with slight modifications to its original code. This important feature enables current users of NPAIRS, i.e. bio-medical scientists, to utilize GPUs without applying fundamental changes to their application. Our experiments show a 7-fold speedup for NPAIRS. Then, we investigated several scheduling algorithms for a heterogeneous CPU-GPU cluster. We show that scheduling can highly improve makespan of a CPU-GPU cluster.Finally, we propose a dynamic scheduling that estimations task execution times on various resource types. Our scheduling scheme obtains performance samples on-the-fly and distributes the tasks based on the historical execution data.