Random Forest predicts HPC configs from just 54 benchmark runs
ML surrogate cuts expensive cluster benchmarking by 95%...
A team led by Dennis Alves Pedersen from CENAPAD-SP trained a cold-start Random Forest surrogate on just 54 structured benchmark runs of LAMMPS+SPICA simulating the antimicrobial peptide Tritrpticin on a hydrated DOPC bilayer (4,354 coarse-grained beads). The benchmarks spanned 18 hybrid MPI+OpenMP configurations across 1-8 AMD EPYC 7662 nodes on the Lovelace cluster, with three independent replications each. Using nine topology and resource features, the model predicted loop time and four internal LAMMPS timing fractions (Pair, Kspace, Comm, Modify) with a mean absolute error of 0.49 seconds (4.0% relative) on loop time.
Feature importance analysis showed that predictive signal is concentrated in OpenMP thread count and the MPI/OpenMP ratio, while raw node and core counts contributed under 3%. Leave-one-dimension-out generalization tests revealed that accuracy degrades when the surrogate crosses hardware boundaries (e.g., from single-node to multi-node, or across threading tiers). Within a common hardware regime, the model ranks configurations correctly, enabling trustworthy recommendations. This approach offers an interpretable map of where surrogate predictions are reliable, allowing HPC teams to reduce costly exhaustive benchmarking campaigns significantly.
- Random Forest trained on 54 runs predicts optimal MPI+OpenMP configs for LAMMPS+SPICA with 4.0% relative error on loop time.
- Feature importance: OpenMP threads and MPI/OpenMP ratio dominate; raw node/core counts contribute less than 3%.
- Model generalizes well within hardware regimes but degrades when crossing between single-node, multi-node, or shared threading tiers.
Why It Matters
Slash HPC benchmarking costs—a cheap ML model can replace thousands of expensive cluster runs for molecular dynamics workloads.