LLNL researchers automate HPC kernel similarity with new metrics
New similarity metrics automatically match benchmarks to real simulation codes.
Researchers from Lawrence Livermore National Laboratory (LLNL) have developed automated performance similarity metrics to evaluate how well benchmarks and proxy applications represent real high-performance computing (HPC) codes. Published on arXiv and set to appear at ICS Workshops 2026, the work by Michael McKinsey, Stephanie Brink, and Olga Pearce addresses a long-standing challenge: manually determining whether a given benchmark accurately reflects the performance characteristics of a simulation code is labor-intensive and hard to scale. Their approach defines two broad categories of computational kernels with similar hardware utilization patterns, then computes pairwise similarity scores.
To test their metrics, the team used the Kripke proxy application and the RAJA Performance Suite, running on both a CPU-only system and a GPU-accelerated system. They successfully validated that their method correctly matches a kernel from Kripke to a counterpart in RAJA. This automated pipeline enables rapid assessment of emerging hardware architectures, helping code developers and hardware designers identify representative benchmarks without manual effort. The work also informs future HPC system design by providing objective similarity scores that can be applied across diverse codebases.
- Two broad categories of kernels are defined based on compute hardware usage patterns.
- Metrics validated on Kripke proxy app and RAJA Performance Suite across CPU and GPU systems.
- Automates the previously manual process of matching benchmarks to simulation codes for scalability.
Why It Matters
Automating benchmark representativeness saves time and helps optimize HPC code for evolving CPU/GPU architectures.