This event is a part of the "Best Practices for HPC Software Developers" webinar series, produced by the IDEAS Productivity Project. The HPC Best Practices webinars address issues faced by developers of computational science and engineering (CSE) software on high-performance computers (HPC) and occur approximately monthly.
| Resource Information | Details |
|---|---|
| Webinar Title | Basic Performance Analysis and Optimization – An Ant Farm Approach |
| Date and Time | 2016-08-09 01:00 pm EDT |
| Presenter | Jack Deslippe (Lawrence Berkeley National Laboratory) |
| Registration, Information, and Archives | https://ideas-productivity.org/resources/series/hpc-best-practices-webinars/#webinar007 |
Webinars are free and open to the public, but advance registration is required through the Event website. Archives (recording, slides, Q&A) will be posted at the same link soon after the event.
Abstract
How is optimizing HPC applications like an Ant Farm? Attend this presentation to find out. We’ll discuss the basic concepts around optimizing code for the HPC systems of today and tomorrow. These systems require codes to effectively exploit both parallelism between nodes and an ever growing amount of parallelism on-node. We’ll discuss profiling strategies, tools (for profiling and debugging) and common issues with both internode communication and on-node parallelism. We will give an overview of traditional optimizations areas in HPC applications like parallel IO and MPI strong and weak scaling as well as topics relevant for modern GPU and many-core systems like threading, SIMD/AVX, SIMT and effectively using cache and memory hierarchies. The “Ant Farm” approach places a heavy emphasis on the roofline performance model and encouraging users to understand the compute, bandwidth and latency sensitivity of their applications and kernels through a series of easy to perform experiments and an easy to follow flow chart. Finally, we’ll discuss what we expect to change in the optimization process as we move towards exascale computers.
Presenter Bio
Jack Deslippe is the application performance group lead at NERSC. Jack and his group are partnering with DOE application teams to evaluate and improve the performance of applications on Cori and future systems at NERSC. He received a Ph.D. from UC Berkeley in physics in 2011, with research centered on computational materials physics and nano-science, including the development and scaling of electronic-structure codes. Jack has been at NERSC since 2011, acting as a consultant and developer for materials science applications and currently leads the NERSC Exascale Science Applications Program (NESAP).
