NEST scaling study

A worked strong and weak scaling study of a Brunel balanced random network in NEST, expressed entirely as labflow experiments. Source: examples/scaling/.

One experiment (brn_scaling) builds and simulates the network and records build + simulate wall-clock times. You choose the regime by how you sweep it.

Strong scaling — fix the problem, add cores

labflow sweep brn_scaling n_total_fixed=50000 threads=1,2,4,8,16,24,48

Total network size is constant; only the thread count changes. Speedup is T(1) / T(p); parallel efficiency is speedup / p.

Weak scaling — grow the problem with the cores

labflow sweep brn_scaling n_per_thread=2500 threads=1,2,4,8,16,24,48

Total size grows with threads (N = n_per_thread × threads), so the work per core is constant. Weak efficiency is T(1) / T(p) — ideally flat at 1.

On a cluster

labflow submit brn_scaling --system cluster-arm --partition normal \
    --nodes 1 --time 1:00:00 --account <account> -- n_total_fixed=50000 threads=48

Each run writes a manifest.json with the returned timings and full provenance (git SHA, config, seed, hardware, wall time).

Analyze

python examples/scaling/analyze.py outputs/brn_scaling --out scaling.png

The analyzer reads every run manifest, auto-detects strong vs. weak, prints a table (build / sim / total time, speedup, efficiency) and saves speedup + efficiency curves.

Extending to multi-node

Single-node OpenMP thread scaling is the simplest study. For multi-node MPI scaling, build NEST --with-mpi, give your system's SLURM template an MPI launcher (srun/mpirun), and submit with --nodes N.