Towards a Scalable Framework for Performance Modeling and Automatic System Optimization --- from many-core to heterogeneous exascale systems

Event Sponsor: 
Mathematics and Computer Science Division
Start Date: 
Feb 25 2010 - 10:30am to 11:30am
Building/Room: 
Building 240, Conference Room 4301
Location: 
Argonne National Laboratory
Speaker(s): 
Jiayuan Meng
Speaker(s) Title: 
PhD Student, University of Virginia
Host: 
Rob Ross

As projected in 2018, an exascale system will have millions of processors with some processors potentially having hundreds of threads of execution. Such large scale, heterogeneous systems challenge system configurations and application adaptations. A potential solution is to establish performance models to quickly explore the space of system configurations and identify application-specific adaptations. Such a framework involves data collection, analysis, modeling, system reconfigurations, and choices of optimizations. It also requires a programming model that allows the same code to adapt to different workload decomposition and distribution schemes. Similar issues already arise in many-core systems, although at a much smaller scale. I will present my dissertation work with a focus on performance modeling and automatic optimization on GPUs, as well as hierarchical task decomposition for many-core architectures. I will then conclude the lessons learned, and talk about the possibilities to extend the work to exascale systems.

Bio:

Jiayuan Meng is a Ph.D. candidate at University of Virginia. His research interests mainly include parallel computation for data-intensive applications and infrastructural design for scalable computation platforms. His dissertation topic is data management for multi-threaded cores. He has built MV5, an event-driven, cycle-accurate many-core simulator based on M5. He has proposed techniques to improve performance scalability on many-core architectures from the aspects of caching protocols, run-time scheduling, dynamically adaptive SIMD architectures, and analytical performance models for GPU applications. He is also collaborating with NEC Laboratories America on domain-specific programming models for emerging recognition and mining applications. He has experiences with various applications including fluid dynamics, image synthesis, 3D rendering, and semantic analysis. He has received the 2009-2010 NVIDIA research fellowship and the 2010 U.Va. Award for Excellence in Scholarship in the Sciences & Engineering.

Miscellaneous Information: