abench

Run reproducible ActivitySim runtime and memory experiments in Linux Docker, from macOS or Linux. One runner supports MTC, SANDAG ABM3, and other models through small YAML profiles. ActivitySim itself does not need to be installed on the host.

With uv, run without managing a Python environment:

uvx abench --help
uvx abench experiments.yaml

For a specific release use uvx abench@0.1.0 experiments.yaml; use uvx abench@latest to refresh to the latest release. Docker and model data must still be available locally. macOS and Linux hosts are supported.

Alternatively, install with pip:

python -m pip install abench
abench run --model-dir /path/to/sandag-abm3-example --profile sandag \
  --source activitysim=ActivitySim/activitysim@<full-40-character-SHA> \
  --source sharrow=ActivitySim/sharrow@<full-40-character-SHA> \
  --multiprocess --processes 4 --sharrow --households 28365 \
  --memory 32g --shm-size 8g --output-dir /path/to/experiments/sandag

Use --profile mtc for MTC; both profiles ship with the package. SANDAG defaults to its small benchmarking-data, not full-scale skims. MTC defaults to data_full. --data-dir overrides either. Model directories need not be Git repositories; Git revision/status are recorded where available and model files are always snapshotted. Existing example scripts and normal configs are untouched.

The host needs Python 3.10+, PyYAML (installed with abench), and a Linux Docker engine with cgroup v2 and memory.peak. Docker Desktop must have enough VM RAM for the chosen memory limit plus VM overhead. The default container is Debian Bookworm/Python 3.11. Current instrumentation requires ActivitySim's workflow.State API (1.4-era or newer); arbitrary historical revisions are not promised to work. Build/runtime failures retain diagnostics and a failure report.

Named experiment files

Write common options once and override only what differs between runs:

schema_version: 1
vars:
  households: 28365
  warmup_households: 5000
  model: /path/to/sandag-abm3-example
output_root: ./results/sandag-${timestamp}
defaults:
  model_dir: ${model}
  profile: sandag
  data_dir: ${model}/benchmarking-data
  config_overlay: ["${model}/configs_explicit_chunk"]
  multiprocess: true
  processes: 4
  sharrow: true
  households: ${households}
  warmup_households: ${warmup_households}
  memory: 80g
  shm_size: 8g
  sources:
    - sharrow=ActivitySim/sharrow@fc175b27d8e0c5d202721c67d96b050e6117b235
runs:
  main:
    sources:
      - activitysim=ActivitySim/activitysim@5c6fae24a91a57a2d6dfc2e1dbe062a61d94545a
  pr1110:
    sources:
      - activitysim=ActivitySim/activitysim@51e298a84276813946e1d623c9a5785e078e022f

Save it as sandag.yaml, then run:

abench sandag.yaml
# Or check all runs without building images or running models:
abench validate sandag.yaml

A ready-to-use SANDAG chunked suite is included in the repository. Its paths assume abench and the SANDAG repository are siblings. The pinned main revision is the one used in the earlier trials, not a moving branch reference.

• defaults accepts CLI options using underscores (shm_size, config_overlay, etc.). Use multiprocess: false for serial execution and sharrow: false to disable Sharrow. sources accepts the same strings/mappings as model profiles.
• runs is an ordered mapping of names to overrides. Each run inherits defaults; ordinary values and lists are replaced. Sources merge by normalized package name, so changing ActivitySim does not discard the shared Sharrow pin.
• ${name} substitutes a reusable scalar from vars; terms can reference other terms. A whole-value reference preserves its type, including numbers/booleans. Undefined references and cycles are errors. No shell or environment expansion is performed. ${timestamp} is a built-in UTC launch identifier shared by all runs, with microseconds to avoid reusing output directories.
• All explicit paths in the suite are relative to the YAML file, independent of the terminal's current directory. This includes overlays and custom profile paths. Built-in mtc/sandag profile names retain their meaning. When omitted, model_dir defaults to the YAML file's directory; the model profile still supplies its usual default data/config paths.
• The suite owns output locations: output_root/<run-name>/. Set output_root once instead of output_dir in each run. Existing roots are rejected.
• All runs are preflighted before the first starts, then run sequentially in file order. Failure stops the suite and retains partial results. The combined report is output_root/comparison.html; individual runs retain their own reports. experiments.yaml and suite.json record the original file and expanded plan.
• File invocations do not accept additional CLI overrides. Edit defaults or the relevant run to keep the file a complete description of the experiment.

This experiment file describes which tests to run. A model profile such as benchmark.yaml describes how to configure a model, and remains reusable across suites.

Run controls

• --single-process (default), or --multiprocess --processes N. The count applies to every sliced stage; coordinators are additional processes.
• --sharrow (default) or --no-sharrow. Sharrow enabled requires its source pin.
• --households N (default 1,000); zero uses the original full input population. abench never replicates households. Positive samples must match realized output.
• --config-overlay configs_explicit_chunk adds config directories in listed priority order. Relative overlay paths are relative to the model directory.
• --memory 16g, --shm-size 8g, --interval 0.5, and optional --platform.
• --output-dir must be new. --label names an experiment, and --compare accepts earlier experiment directories. Compatible compiled flows are reused automatically; the serial warmup still runs.

abench validate accepts the same experiment arguments without --output-dir. It checks the profile, required inputs, CSV population size, source pin syntax, and Docker capabilities without building an image or running the model. It does not prove Git commit availability, package compatibility, or skim consistency; those are checked by the build and model run.

Any dependency from GitHub source

Repeat --source for any Python distribution, including add-on extensions:

--source 'my-addon[fast]=ExampleOrg/model-addon@<SHA>#subdirectory=python/addon'

The left side is the distribution name (which can differ from its import module). Each source uses an exact full SHA and an organization/repository name. Extras and a repository subdirectory are optional. Profiles can declare the same entries as strings or mappings:

sources:
  - name: my-addon
    repository: ExampleOrg/model-addon
    commit: '0123456789abcdef0123456789abcdef01234567'
    extras: [fast]
    subdirectory: python/addon

CLI sources override profile sources by normalized distribution name. Duplicate CLI entries are errors. --activitysim-commit and --sharrow-commit remain aliases for the official repositories; conflicting alias/source declarations are errors.

Inside Docker, abench verifies each checkout's Git object, builds a wheel, checks its distribution name, then installs all source wheels together with other requirements. It runs pip check and verifies installed wheel identities. Exact source commits, resolved versions, and the full dependency environment are saved. All source dependencies must agree: conflicting requirements fail the build.

Use profile requirements for additional registry requirements and constraints for resolver bounds or exact transitive pins. Profiles may select python_image (a compatible Debian-based image, optionally pinned by digest). The default image includes a compiler and HDF5 headers. Packages requiring other system libraries can use a prebuilt compatible base image. Private GitHub authentication and custom OS provisioning are outside the initial interface.

Source pins do not freeze unpinned transitive/build dependencies or base images. Retain images and dependency manifests for strict reproduction. Build isolation may fetch build requirements; constraints currently govern the final environment, not those isolated build environments.

Add another model

Create benchmark.yaml in its model directory, then use --model-dir:

schema_version: 1
name: My regional model
configs: [configs]
mp_configs: [configs_mp]
snapshot: [configs, configs_mp, extensions]
extensions: [extensions]
data_dir: data
required_inputs:
  - [households.csv, households.parquet]
  - persons.csv
  - land_use.csv
  - skims.omx
settings:
  use_shadow_pricing: false
  rng_base_seed: 0
input_tables:
  households: {}
  persons: {}
  land_use:
    totals: [TOTPOP, TOTHH, TOTEMP]
    zone_columns: [TAZ]
output_tables:
  households: {}
  persons: {}
  tours: {}
  trips:
    categories: [trip_mode, primary_purpose]

Paths in a profile are relative to the model directory (except data_dir, which may be absolute). configs and mp_configs are ordered, highest priority first. snapshot must cover config files, local extension modules, and adapter modules; entries must not overlap or contain directory symlinks. Data is mounted read-only and should not change during a run. Profiles are trusted model code/configuration.

Settings precedence, lowest to highest: normal model configs → profile settings → user overlays → required CLI controls (sample, SP/MP, worker counts, Sharrow, and fail-fast). Generated inheriting profile configs preserve this ordering when ActivitySim reconstructs worker settings. Component overlays remain independent.

Optional profile fields:

For specialized data formats, an adapter can arrange compatible CSV/Parquet summary outputs. The initial generic reader does not interpret arbitrary binary model outputs.

Measurement and reports

Sharrow runs first execute the model in a separate single-process warmup, using min(target households, 5000) households by default. For --households 0, the target is the full available population, so warmup uses at most 5000 of those households. Set --warmup-households N (or warmup_households: N in experiment YAML) to change this positive cap. Warmup always uses one process; measured runs retain their requested sample and worker count. Model config directories, seed, chunk overlays, and flow cache path are retained from the target experiment.

Compiled flows are automatically reused across runs and suites, including changes in ActivitySim revisions, sample sizes, process counts, and model configs. The persistent host cache defaults to ~/.cache/abench/flows. Change it with --flow-cache-dir PATH (flow_cache_dir in YAML), or disable automatic reads and writes with --no-reuse-flows (reuse_flows: false). --cache-from remains an explicit seed option with its existing stricter dependency checks.

Compatibility uses the installed Sharrow, Numba, llvmlite, and NumPy versions, plus their source repository/commit identities when applicable, Python version, and container architecture/CPU features. ActivitySim and model settings are excluded from this key: Sharrow identifies generated flows by their contents, and Numba checks cached signatures. Changed flows can compile during warmup. An existing cache does not guarantee that warmup will need no compilation.

Each experiment receives a private copy with source timestamps preserved. Warmup always runs, and successful warmups atomically update the persistent cache before measurement starts. Compatible simultaneous experiments wait for each other to avoid losing compiled signatures when attempts update Numba cache indexes. Only cache/flows is shared, never model data, outputs, or shared-memory artifacts. Cache identity and reuse counts are recorded in flow-cache-identity.json and experiment.json. The persistent cache can be deleted between runs to reclaim disk space; older experiments created before this feature are not automatically imported.

A smaller serial warmup may not exercise every flow/type signature needed by the measured run. Each measured attempt therefore records flow compilation and allows it to finish. If compilation occurred, the completed attempt becomes cache preparation, and none of its runtime or memory results qualify as benchmark results. Its diagnostics and outputs are retained under attempts/attempt-001, attempts/attempt-002, etc. Newly compiled flows are published to the shared cache.

The model then restarts in a fresh container with fresh outputs and model caches, using the same settings and expanded flow cache. Only an attempt with zero flow compilations is accepted. All attempts keep permanent directories under attempts/; measured/ links to the accepted attempt. By default abench allows two additional attempts (three total). Set --cache-retries N or cache_retries: N in YAML; zero allows no retries. If compilation persists, the experiment fails with the final attempt's diagnostics retained. Ordinary model errors, OOMs, and output validation failures stop immediately and are never retried as cache preparation.

The report and experiment.json include attempt history. Each attempt retains cache-miss-details-*.jsonl, its settings, component timings, and memory samples. Ordinary non-flow compilation and disk-cache loading remain included in accepted measurements. Flow tracking uses private Numba internals and is covered by real compilation, cache-hit, and Docker retry tests.

Memory is the whole-container cgroup v2 charge, counting shared pages once. Blue is memory.current (including file cache, shared memory, and kernel costs). Green dashed is anon + shmem, a subset excluding ordinary file cache and kernel costs. Never add the lines. Swap is recorded separately and disabled by equal memory/memory+swap limits. /dev/shm capacity is within that limit.

Component runtimes show worker mean, population SD, count, and maximum. Worker SD is imbalance, not confidence across repeated trials. The component dropdown highlights each worker's actual window; overlaps darken and gaps remain clear. Overall elapsed includes startup, coordination, and checkpoint writes between components. Kernel peak includes startup; warmup and post-run summaries are excluded. Shared VM page-cache ownership can influence container charges: this is not a cold-input I/O benchmark.

abench report --compare /path/to/run-a /path/to/run-b \
  --output-dir /path/to/comparison.html

Reports are offline HTML/SVG/JavaScript plus normalized JSON. They retain fastest successful component highlighting and use common memory axes. Existing MTC and SANDAG schema-1 experiments remain readable, including approximate legacy timing windows where only completion logs exist. New experiments use schema 2 and include source manifests, resolved profile/settings, model and harness snapshots, file hashes, Docker details, and source installation provenance. Input file size/mtime records are provenance hints, not content hashes of large skims.

Development and CI

python -m pip install -e '.[test]' -r tests/requirements.txt
python -m pytest -m 'not docker'
pre-commit run --all-files
python -m build
ABENCH_DOCKER_TESTS=1 python -m pytest tests/test_docker.py -v

Install Node.js to exercise the offline chart selector test. GitHub Actions runs unit tests on Python 3.10–3.12, lint/format checks, wheel packaging checks, and Linux Docker integration. The Docker tests build pinned ActivitySim/Sharrow sources and run a four-household extension workflow in serial and multiprocess modes, checking warmup, generated-code cache reuse, per-worker timings, merged outputs, memory, and reports. They do not require either example repository or large datasets.

Adapted from the MTC and SANDAG benchmark harnesses developed in this workspace. The original measurement approach was informed by WSP's Lighthouse production benchmark. See LICENSE for the retained BSD license.

Releases

See RELEASING.md for Trusted Publishing setup and release instructions.