Stanford High Performance Computing Seminar Series
There are many computer science classes that provide comprehensive understanding of computer science theory, from algorithms to artificial intelligence. However, classes that prepare students to use advanced computing resources as they are used in computational, applications-driven research and development are relatively rare in university curricula. This seminar series provides an introduction to the use of advanced computing resources with real-world examples of large-scale, multidisciplinary, simulation-based science as related to academic and applied research.
Predictive Parallel Performance Models for Petascale Platforms
Monday May 18th at 12:00 pm
Curtis W. Hamman
Center for Turbulence Research
In this talk, I will show how you can take advantage of predictive performance modeling to identify bottlenecks in your own codes and discover new strategies for parallelism on petascale platforms. The historical development of performance modeling from Amdahl's Law in 1967 to the modern theory behind performance engineering on heterogeneous architectures will be highlighted. Methods to model and quantify the interplay between computation, data movement, cache coherency, synchronization, and I/O will also be discussed. Practical examples using the modern suite of SciDAC-developed tools for performance modeling and analysis will be given. Several case studies will be analyzed with a specific focus on predicting the performance and impact of algorithmic design choices in the direct numerical simulation of turbulent channel flows.
Be sure to join us next Monday June 15th 12:00 PM at Wallenberg Theater
(Wallenberg Hall, Building 160, Main Quad)
Previouse Seminar Series: