Egan Johnson

Welcome

I am a PhD student studying compilers for tensor programs and distributed systems at the University of Illinois Urbana-Champaign. I am advised by Professors William Moses and Sasa Misailovic.

Interests

As tensor programs scale ever larger, driven by increased demand for large machine learning and scientific workloads, new challenges and developments emerge all across the system stack. I am interested in:

  • Distributed Compilers: Machine learning and scientific programs are often executed on handfuls to hundreds of distinct devices, and managing costs and overheads introduced by distribution is key to achieving high performance and energy effiency. However, distribution strategies are often made manually by programmers. I am interested in automatic compiler reasoning and transformation of distributed programs.
  • Tensor Compilers: Even in the context of a single device, tensor programs require a new and evolving set of optimizations to bring down their computational costs. Many tensor programs involve a high degree of manual tuning or kernel development: I am interested in how compilers can be applied to achieve similar performance automatically.
  • Compilers for Emerging Hardware: As tensor programs grow more expensive, hardware developers rise to meet the challenge. However, accelerators require finely tuned software to achieve full utilization. A full accelerator programming stack requires a strong compiler.

Posts

No posts yet.