The ACM International Symposium on High-Performance Parallel and Distributed Computing (HPDC) is an academic research conference, started in 1992, that focuses on topics around systems software and middleware layer of high-performance and distributed computing. It’s historically been attended by academic researchers (CS faculty, grad students) with some DOE research attendees and limited industry presence. The conference has historically leaned towards job scheduling, workflow management, and distributed runtimes.
I was invited to HPDC 2026 in Cleveland, OH for two reasons:
- Present the opening talk at the PERMAVOST workshop
- Contribute a short talk to the Moonshot Ideas session
PERMAVOST
I will be presenting the following:
Title: Workflow performance is not the sum of its parts
Abstract:
Performance and progress have long been synonymous in the world of high-performance computing: a faster system was a better system, so performance tools and benchmarks were an effective way to predict how productive a supercomputer could be. This held because models were fixed, and computers only had to execute them quickly. Data-driven discovery (and AI) breaks this equivalency, because models must be trained and refined while they produce output. A machine can run at full speed while making no progress at all, because today’s performance tools measure speed, not progress.
This forces a change in how performance must be measured. Because their models evolve over time, AI workflows have a far more complex relationship with infrastructure than traditional simulation, one that must be understood end-to-end, not layer by layer. Using production examples, I will show how individual components can each perform optimally in isolation while the composed system still runs data-driven workflows inefficiently. Performance is not the sum of its parts, and I will offer a few directions in which performance analysis tooling might need to evolve to capture performance as it actually emerges from these interactions.
Moonshot Ideas session
Taking the Seatbelts Out of the Supercomputer
In supercomputing, the biggest challenges often involve getting people to change their behavior. What happens when people are taken out of the picture, and agents become the predominant users of supercomputers? What decades-long methods, software stacks, and fundamental architectures exist solely for human convenience, and what optimizations are possible in a parallel computing stack that is not user-friendly at all? In this talk, I will explore this idea through a few specific cases where technologies we take for granted in HPC are actually attractive nuisances in an environment used by agents.