In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy

Jeremy Kemp

will defend his dissertation proposal

Extending Directive Based Tasking Models to Exascale

Abstract

With every major shift in high performance computing hardware, new programming and execution models arise to leverage them. First vector processors, then communication libraries, both single sided and two sided emerged to be the default means of expressing parallelism for super computers. As multicore processors became the default, new APIs emerged and old ones adapted to the new hardware. As GPGPUs and other accelerators emerged and increased in popularity, again, APIs were ma de and extended to accomodate them. Now, as exescale nears, it becomes obvious that not only are all of these tools going to be needed, but they need to be coordinated with the most precise synchronization possible.

Previous generations of applications have been able to rely on barriers and other heavy synchronzations, but as the size of a node scales up, this becomes less feasible. A new generation of distributed tasking APIs is being developed to address these problems using lightweight threads and data dependencies as the primary tools. However, moving existing applications to these new models is difficult to the point where a complete rewrite is often advisable. In this work, we propose API extensions and an implementation for OpenMP that would enable hybrid applications, a path to port applications incrementally, and an interface as flexible as HPX, the asynchronous task based programming model that the extensions are built on, or any other of the emerging exascale models.

Date: Monday, November 30, 2015
Time: 10:00 AM
Place: PGH 218
Advisor: Prof. Barbara Chapman

Faculty, students, and the general public are invited.