Student research opportunities

Diamond Tiling in 3D for increasing parallel scalability of stencil computations

Project Code: CECS_941

This project is available at the following levels:
CS single semester, Honours, Summer Scholar, Masters

Keywords:

compilers, polyhedral model, tiling, parallel programming, scientific computing, high performance computing

Supervisor:

Dr Michelle Strout

Outline:

Stencil computations are used anytime partial differential equations are being solved on regular grids; combustion simulations and atmospheric science simulations are two example applications we work with in our research group. The main issue with these computations is that they hit a memory bandwidth performance bottleneck when they are being executed in parallel. In collaboration with Dave Wonnacott at Haverford and Uday Bondhugula at the Indian Institute of Science, we are experimenting with a new way to schedule such stencil computations that improves the data locality of the computation by grouping computations into 3D "diamonds". and feel free to send email to Dr. Michelle Strout at mstrout@cs.colostate.edu to set up a meeting for more details. Michelle is at ANU from July 23, 2013 through May 31, 2014 for sabbatical.

Goals of this project

We are currently running experiments to determine the scalability of this approach, the loop overhead, and working on ways to improve and automate visualization.

Requirements/Prerequisites

A student working on this project should be comfortable programming in C++ and have some parallel programming experience.

Student Gain

This work will most probably lead to a publication. Additionally there are funded PhD positions at Colorado State University for students who excel at this project and whom apply for graduate school at CSU.