My PhD thesis focuses on exploring the possibilies of using the paradigm of vector processing in a many core environment as a way of increasing system performance. We consider the vector unit as an accelerator managed by one or more CPUs running the operating system and runtime libraries.

We split vector computation tasks into smaller parallel subtasks that are distributed to different accelerators. The distribution of tasks can be done by either moving data to computation cores or moving tasks to cores closer to the data. Our goal is to quantitatively determine how overall performance scales when increasing the number and size of the vector accelerators. We also analyze the various scheduling methods and asses their performance.

To test and evaluate our proposed techniques we are developing a cycle accurate simulator SARCSim which is built on the Unisim infrastructure.