Accelerators and Hybrid Exascale Systems (AsHES)
IPDPS 2013: IEEE International Parallel and Distributed Processing Symposium
Prof. Richard Vuduc, Georgia Institute of Technology
In this talk, we will "pop up a level" and ask whether first principles of algorithm and architecture design can tell us anything about heterogeneity. In particular, I'll posit "strawman" cost models that connect properties of an algorithm with physical properties of machines, like energy, power, and area. I will then explain what these models imply. For instance, I'll use them to evaluate the time- and energy-efficiency of heterogeneous designs as well as for predicting the benefits of other forms of heterogeneity, such as "communication" heterogeneity. And although I'll show a bunch of formulas, calculate things, and plot them, I stress that this talk is about ideas, rather than a well-developed set of results. As such, your questions, healthy skepticism, (constructive!) feedback, and offers of collaboration may be even more welcome than usual! :-)
This work is joint with current and former students: Jee Whan Choi, Aparna Chandramowlishwaran (MIT), Marat Dukhan, and Kenneth Czechowski. The talk will preview two papers (one lead by Choi, the other by Czechowski) to appear at the main IPDPS'13 conference later in the week.Bio:
Richard (Rich) Vuduc is an Associate Professor at Georgia Tech in the School of Computational Science and Engineering. His research lab, the HPC Garage (hpcgarage.org), is interested in all-things-high-performance-computing, with an emphasis on parallel algorithms, performance analysis, and performance tuning. He is a member of the DARPA Computer Science Study Panel, recipient of the NSF CAREER Award, and co-recipient of the Gordon Bell Prize (2010). His lab's work has received a number of best paper nominations and awards, including most recently the 2012 Best Paper Award from the SIAM Conference on Data Mining.
ASHES steering committee