Mihail Popov (Storm)

Greetings, everyone!

Winter is coming, and so are our seminars. Thankfully, the seminars are arriving before winter.

Tune in next Friday to learn more about Mihail Popov (et al.) work on design space exploration and optimization using his machine learning methods.

You can join us on September 30 (Friday) at 2 pm in room George Boole 2 or online.

Speaker

Mihail Popov (Storm)

When

Friday, September 30, 2022 at 2 pm.

Where

room George Boole 2 or online (Zoom link below)

Title

Characterizing Design Spaces Optimizations

Abstract

High performance computing systems are increasingly difficult to use. They are massively parallel, heterogeneous, with NUMA effects, out-of-order execution, cache hierarchies, and data prefetching. The resulting complexity requires adapting applications to the hardware characteristics to achieve full performance and energy efficiency. The software stack provides many diverse knobs to adapt an application to the hardware. At compilation, different optimization passes can be selected. At execution, the runtime provides an interface to control parallelism, thread placement, data placement, and workload scheduling. Interestingly, hardware capabilities can also adjust the system to the application (e,g, prefetch, cache, frequency). Therefore, achieving efficiency requires exploring a huge space of hardware and software optimizations.

This talk focuses on exploring and describing such software and hardware optimizations with machine learning.

In particular, it will present insights from our last 3 papers done in collaboration with: Amina, Emmanuelle, Laercio, Lana, and Olivier as well as several other external collaborators.

I am also quite interested to discuss future ideas :)