Nowadays, general-purpose on graphics processing units (GPGPUs) show high performance on accelerating the data parallel computing in many domains due to its powerful parallel processing capabilities and programmable pipelines, especially for problems with regular structures. There are, however, many irregular problems which are difficult to predict how long a task will take to complete and how many new tasks are created dynamically during runtime, like hierarchical matrix construction and arithmetic, PageRank. It is difficult to provide an attractive advantage in performance when porting these irregular problems to GPU straightforwardly, because GPUs depend on uniform work distribution to expose their full potential with data parallel designs. To solve problems of the difficulty of task scheduling in irregular applications on GPU systems, I would like to propose a new task scheduler using task parallelism, where a dynamic load balancing among GPUs is achieved using work stealing strategy. I will also integrate the task scheduler for GPU into a task parallel language, Tascell, to make it can deal with GPU computing in a relatively small programming effort. I will apply this proposal to several irregular applications, like hierarchical matrices (H-matrices) construction and arithmetic and evaluate the performance.
High Performance Computing, GPGPU, Task Parallelism
| 氏名 | コース | 研究室 | 役職/学年 |
|---|---|---|---|
| Jing Xu | システム科学コース | 岩下研究室 | 博士2回生 |