Co-Modeling Workload and Hardware for the Next Generation of Heterogeneous Computing
|Event Date:||February 28, 2014|
|Speaker:||Dr. Jiayuan Meng|
|Speaker Affiliation:||Argonne National Laboratory|
|Contact Name:||Prof Anand Raghunathan
|Contact Phone:||(765) 494-3470
|Open To:||ACCEPTABLE FOR ECE694
Scientific applications such as climate modeling and cosmology simulation often have a longer lifespan than Microsoft Office and yet they have to be executed efficiently on the most recent supercomputers. How to explore better implementations on emerging systems? And if you are a hardware architecture, how would you design your system to expect software evolution? The growing complexity in applications and system architectures has already increased the knowledge gap among application developers, performance engineers, and hardware designers. We need a system to deal with complex systems.
Towards this goal, we propose a model-based framework, SKOPE, that produces a descriptive model about the semantic behavior of a workload, which, when combined with hardware models, can infer potential transformations and help users understand how workloads may interact with and adapt to emerging hardware. In particular, this talk will focus on a GPU auto-tuning framework named GROPHECY, which uses SKOPE to identify high-potential transformations and estimate the performance outcome. It also has been used to survey hypothetical GPU designs.
Jiayuan Meng is an Assistant Computer Scientist at Argonne National Laboratory. He has been researching on heterogeneous system infrastructures for scientific applications, with a specific focus on performance modeling, optimization, and software-hardware co-design. He earned his Ph.D. in Computer Science from University of Virginia in 2010. His MV5 simulator is used by various institutions to study heterogenous many core architectures. He received the 2009-2010 NVIDIA Graduate Fellowship and the 2010 U.Va. Award for Excellence in Scholarship in the Sciences and Engineering.