NetLab

NetLab builds network scenarios with NetLab-TopoGen, runs them with NetGraph, and compares the results. It reports how much traffic survives failures, how path costs change, and how much the modeled equipment costs and consumes.

You can run several configurations and random seeds in one batch. NetLab also supports parameter searches and experiments proposed by a language model.

Install

Use Python 3.11 or later in a virtual environment of its own. To install the command:

python -m pip install netlab

The command also installs NetGraph (ngraph), NetLab-TopoGen (netlab-topogen) and the other dependencies from PyPI; pyproject.toml lists their supported versions.

ipSpace's networklab package installs a command of the same name, so use separate environments for the two tools.

To use the included examples, clone the repository with Git and create its environment:

git clone https://github.com/networmix/NetLab
cd NetLab
make dev
source venv/bin/activate

Run the commands below from the repository root.

Run an example

This example uses saved geographic routes, so no Census download is needed. It still needs network access: the build draws a site map for each scenario from downloaded map tiles. To build offline, set visualization.site_graph.export: false in a copy of the configuration.

netlab run topogen_configs_small/small_baseline.yml --seeds 11 12 \
  --graphs-dir tests/data/scenarios/small_baseline/small_baseline \
  --scenarios-dir scenarios
netlab metrics scenarios/

The first command builds a network for each seed and runs its NetGraph workflow. A seed is the scenario's random seed. It drives NetLab-TopoGen's random traffic choices and NetGraph's failure sampling, so the same seed repeats the same experiment. The second command calculates metrics and renders comparison figures.

To build without simulating, use netlab build with the same input arguments. To generate geography yourself, omit --graphs-dir and first download the data required by NetLab-TopoGen. The data/ and lib/ paths are relative to your working directory. From the repository root, lib/workflows.yml replaces NetLab-TopoGen's design_analysis_brief workflow with NetLab's version, and lib/failure_policies.yml adds mc_hard.

Saved geography must be a corridor MultiGraph: a graph that can keep several routes between the same metro pair. Each file is named <config>_integrated_graph.json. The topogen_configs/ directory contains larger designs; topogen_configs_small/ contains smaller versions for experimentation.

Control a batch

Time limits start when a task starts running. Scenario assembly runs every time. NetLab reuses saved work only when a hash of its inputs still matches; a hash identifies file contents, so any change to the inputs invalidates the cache. Generated geography is reused while the configuration, data files and package versions are unchanged. A simulation result is reused while the scenario, package versions, the saved result file and its run record are unchanged. Failed tasks make the command return a failure status; independent tasks continue. Each run's outcome and the wall time of each workflow step are saved in a run record, <name>.results.run.json, beside the results; a reused result keeps the times of the run that produced it.

Read and compare metrics

netlab metrics scenarios/ --no-plots
netlab metrics scenarios/ --only small_baseline
netlab metrics scenarios/ --summary

--no-plots calculates tables without figures. --only takes a comma-separated list of configuration names, such as small_baseline,small_clos. --summary reads the saved tables and renders figures. A calculation replaces the complete report directory, even with --only. If calculation fails, the existing report is kept.

The workflow must include these step names:

--enable-maxflow also needs a node_to_node_capacity_matrix step for pairwise capacity analysis. Each result needs the run record NetLab writes beside it. See metrics.md for definitions, required fields, worked examples and output files. Failure statistics describe the configured experiment; they are not measurements of network uptime.

To compare two configurations with matching seeds:

netlab test scenarios/ small_baseline small_clos

Both configurations must have been analyzed. The statistical test needs at least three matching finite seed results; the two-seed quick start is for running the pipeline, not statistical inference.

Python API

Save this as a script and point it to a NetGraph scenario:

from pathlib import Path
from netlab.simulation import run_simulation

if __name__ == "__main__":
    outcome = run_simulation(Path("scenario.yml"), timeout=60)
    print(outcome.status, outcome.error)

run_simulation runs in a local worker process. A successful run saves scenario.results.json next to the scenario; every run records its status and the wall time of each step it ran in scenario.results.run.json. The __main__ guard keeps a spawned worker from running the script again. Use simulate(yaml_text) to run in the calling process, or inspect_scenario(yaml_text) to validate and count network objects. See architecture and batch APIs.

Metric functions read a saved result dictionary:

import json
from netlab.metrics.bac import compute_bac
from netlab.metrics.latency import compute_latency_stretch
from netlab.metrics.msd import compute_alpha_star

with open("scenario.results.json") as f:
    results = json.load(f)

print(compute_alpha_star(results).alpha_star)
print(compute_bac(results, step_name="tm_placement").auc_normalized)
print(compute_latency_stretch(results).failures.get("p99"))

They return the per-seed values described in metrics.md. netlab metrics also checks the result arithmetic before it computes them.

Automated experiments

An autoresearch project gives a language model a task description (program.md), the parameters it may change (hypothesis_template.yml) and the metric to optimize (objective.yml). init copies a base scenario that has a MaximumSupportedDemand step and at least one ${{name}} placeholder. Each placeholder becomes a parameter, and the objective maximizes steps.msd_baseline.result.alpha_feasible.value, the multiplier of the step named msd_baseline. A candidate whose search finds no feasible multiplier fails; it never scores zero. Review the generated files before a run:

netlab autoresearch init --base-scenario scenario.yml --output project/
netlab autoresearch run project/ --backend mock --max-experiments 1

The mock backend is for testing the workflow. To use an external model, select claude-cli, codex-cli or openai and configure that backend's credentials. CLI backends need their executable installed and signed in. HypothesisManager can also turn a free-text hypothesis into a scenario, run it and request an interpretation. Those cycle files are saved under <project>/cycles/.

Run one writer per project; nothing locks the files. A project directory is tied to its input files, seed and installed package versions. A sweep output directory is tied to its sweep type, seed, iteration count and package versions. If any of them changes, NetLab refuses to continue there; start a new directory. Rerunning a sweep in its directory skips successful entries and retries failed ones. Higher scores rank better; candidates that fail the configured constraints cannot become best.

For data-center-to-backbone designs, see the research project and included interconnect experiments.

Development

make check-ci

This checks formatting, lint, types and tests, including NetLab-TopoGen → NetGraph → metrics runs using saved geography. See fixture regeneration and metric verification.

In Superset, bash .superset/workspace.sh setup creates a venv, installs the development dependencies and copies missing untracked .env files from the root checkout. The Run button executes make check-ci.

To test local checkouts of all three dependencies:

bash dev/check_ngraph_integration.sh /path/to/NetGraph /path/to/NetGraph-Core /path/to/NetLab-TopoGen

Replace the paths with your checkouts. The script builds Core and runs checks in a temporary environment. It saves source revisions, dependency versions and results under build/ngraph-integration/.

Changelog · MIT license