Optimise Simulation() init: ~10x speedup on warm loads#1497
Merged
nikhilwoodruff merged 4 commits intomainfrom Feb 18, 2026
Merged
Optimise Simulation() init: ~10x speedup on warm loads#1497nikhilwoodruff merged 4 commits intomainfrom
nikhilwoodruff merged 4 commits intomainfrom
Conversation
…rised interpolation Three performance changes that give ~10x speedup on warm Simulation() calls: 1. Cache the fully-processed parameter tree in CountryTaxBenefitSystem.__init__(). convert_to_fiscal_year_parameters() (22,538 param.update() calls, ~0.5s) now only runs once per process. Subsequent inits clone the cached tree. Reforms still go through the full pipeline via apply_parameter_changes() -> reset_parameters() -> process_parameters(), so correctness is preserved. 2. Cache the loaded, uprated and enum-pre-encoded UKMultiYearDataset per URL. _pre_encode_enum_columns() converts string enum columns to int16 before caching so subsequent build_from_multi_year_dataset calls use encode()'s fast integer path. Saves ~2.2s (HDF5 read + uprating + string encoding) on every warm load. 3. Vectorise attends_private_school.interpolate_percentile. Replace the Python list comprehension over ~115k households with a 21-point parameter lookup, np.interp and numpy array indexing. Benchmark: 2nd+ Simulation() drops from ~3.7s to ~0.38s. Co-Authored-By: Claude <noreply@anthropic.com>
nikhilwoodruff
force-pushed
the
perf/brma-lha-optimisation
branch
from
ca4956f to
14dcc79
Compare
Removes the -m "not microsimulation" exclusion from make test, so reform impact and salary sacrifice cap tests run in CI. Updates stale expected values in reforms_config.yaml and test_salary_sacrifice_cap_reform.py to match current model output. Co-Authored-By: Claude <noreply@anthropic.com>
nwoodruff-co
pushed a commit
that referenced
this pull request
Feb 19, 2026
…m labels After the warm-load caching optimisation (#1497), enum columns in dataset DataFrames are stored as int16 for performance. Add decoded_person, decoded_benunit and decoded_household properties to UKSingleYearDataset that return copies of the DataFrames with enum columns decoded back to their string names (e.g. 0 -> 'MALE'). The internal int16 representation is unchanged so simulation performance is unaffected.
4 tasks
nwoodruff-co
pushed a commit
that referenced
this pull request
Feb 19, 2026
…ings Replace plain DataFrame attributes with properties that decode int16 enum columns back to string names on access. The raw int16 data is stored in private _person/_benunit/_household attributes and accessed via .tables by the simulation engine, preserving the warm-load performance gain from #1497. Callers (e.g. sim.dataset[year].person) now see string labels as before.
nwoodruff-co
pushed a commit
that referenced
this pull request
Feb 19, 2026
…ehold (#1498) * Add decoded_person/benunit/household properties to restore string enum labels After the warm-load caching optimisation (#1497), enum columns in dataset DataFrames are stored as int16 for performance. Add decoded_person, decoded_benunit and decoded_household properties to UKSingleYearDataset that return copies of the DataFrames with enum columns decoded back to their string names (e.g. 0 -> 'MALE'). The internal int16 representation is unchanged so simulation performance is unaffected. * Make UKSingleYearDataset.person/benunit/household decode enums to strings Replace plain DataFrame attributes with properties that decode int16 enum columns back to string names on access. The raw int16 data is stored in private _person/_benunit/_household attributes and accessed via .tables by the simulation engine, preserving the warm-load performance gain from #1497. Callers (e.g. sim.dataset[year].person) now see string labels as before. * Format with black * Update UC taper reform expected impact to -44.8bn
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Three changes that together give ~10x speedup on repeated
Simulation()calls (3.7s → 0.38s warm).Parameter tree cache.
CountryTaxBenefitSystem.__init__()clones a pre-processed parameter tree instead of runningconvert_to_fiscal_year_parameters()(22,538param.update()calls, ~0.5s) on every instantiation. The first init populates the cache; subsequent inits clone it in ~0.15s. Reforms still run the full pipeline viaapply_parameter_changes()→reset_parameters()→process_parameters().URL dataset cache. After the first HuggingFace load, the fully-uprated, enum-pre-encoded
UKMultiYearDatasetis cached in memory._pre_encode_enum_columns()converts string enum columns toint16before caching so subsequent loads useencode()'s fast integer path. Saves ~2.2s (HDF5 read + uprating + string encoding) per warm simulation.Vectorised
interpolate_percentile. The Python list comprehension over ~115k households inattends_private_schoolis replaced with a 21-point parameter lookup,np.interp, and numpy array indexing.Correctness verified: enum variables (
region,tenure_type), income tax and household net income all give identical results between cold and warm runs. All tests pass except two pre-existing failures intest_reform_impacts.py.