Implement scalability benchmark sweep

[gregorweiss]

Automated benchmark that discovers optimal resource configuration for a given system size by running short production trials across a sweep of resource configs.

Motivation

Different system sizes have different optimal resource configurations. A 5,000-atom system doesn't benefit from 64 cores the way a 500,000-atom system does. Currently the pipeline uses hardcoded settings (num_cores: 8, num_gpus: 1). A short benchmark sweep identifies the throughput sweet spot before committing to a full campaign.

Depends loosely on: #7, #8 (Cluster autodiscovery — knows available core counts and GPU types)

Scope

• mdfactory/performance/benchmark.py
• BenchmarkConfig dataclass: sweep parameters (core counts, GPU replicas, duration)
• run_benchmark_sweep(system_path, config) -> BenchmarkResult
• GROMACS md.log parser: extract ns/day from performance section
• Result selection: best ns/day or best ns/day per core-hour
• Short benchmark MDP template (100 ps, minimal output)
• Result persistence as JSON sidecar per campaign

Sweep dimensions

• CPU path: vary cores (1, 2, 4, 8, 16, 32, 64) with OMP_NUM_THREADS
• GPU path: vary concurrent replicas (1, 2, 4, 8, 16) sharing one GPU

Reuse across HT batch

If all systems in a campaign have similar atom counts (within ±20%), the benchmark runs once on a representative system. The discovered settings then apply to all production runs in the batch.

Acceptance criteria

• Parser correctly extracts ns/day from GROMACS md.log samples
• Sweep generates correct srun commands for each configuration
• Result JSON contains all trials with ns/day and selects optimum
• Works without cluster (dry-run mode that generates scripts without submitting)