Benchmark utils

Common utilities for benchmark scripts.

This module provides a standardized framework for running benchmarks with:

• Configurable executor types ( ThreadPoolExecutor, ProcessPoolExecutor, InterpreterPoolExecutor)

• Warmup phase to exclude executor initialization overhead

• Statistical analysis with confidence intervals

• CSV export functionality

• Python version and free-threaded ABI detection

See also

• Benchmark tarfile

• Benchmark wav

• Benchmark numpy

Source

Source

Click here to see the source.

#!/usr/bin/env python3
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.

"""Common utilities for benchmark scripts.

This module provides a standardized framework for running benchmarks with:

- Configurable executor types (
  :py:class:`~concurrent.futures.ThreadPoolExecutor`,
  :py:class:`~concurrent.futures.ProcessPoolExecutor`,
  :py:class:`~concurrent.futures.InterpreterPoolExecutor`)
- Warmup phase to exclude executor initialization overhead
- Statistical analysis with confidence intervals
- CSV export functionality
- Python version and free-threaded ABI detection

.. seealso::

   - :doc:`./benchmark_tarfile`
   - :doc:`./benchmark_wav`
   - :doc:`./benchmark_numpy`

"""

__all__ = [
    "BenchmarkRunner",
    "BenchmarkResult",
    "ExecutorType",
    "get_default_result_path",
    "load_results_from_csv",
    "save_results_to_csv",
]

import csv
import os
import sys
import time
from collections.abc import Callable
from concurrent.futures import (
    as_completed,
    Executor,
    ProcessPoolExecutor,
    ThreadPoolExecutor,
)
from dataclasses import asdict, dataclass, field
from datetime import datetime, timezone
from enum import Enum
from functools import partial
from sys import version_info
from typing import Any, Generic, TypeVar

import numpy as np
import psutil
import scipy.stats

T = TypeVar("T")
ConfigT = TypeVar("ConfigT")


def _is_free_threaded() -> bool:
    """Check if Python is running with free-threaded ABI."""
    try:
        return not sys._is_gil_enabled()  # pyre-ignore[16]
    except AttributeError:
        return False


_PYTHON_VERSION = f"{version_info.major}.{version_info.minor}.{version_info.micro}"
_FREE_THREADED = _is_free_threaded()


@dataclass
class BenchmarkResult(Generic[ConfigT]):
    """BenchmarkResult()

    Generic benchmark result containing configuration and performance metrics.

    This class holds both the benchmark-specific configuration and the
    common performance statistics. It is parameterized by the config type,
    which allows each benchmark script to define its own configuration dataclass.
    """

    config: ConfigT
    """Benchmark-specific configuration (e.g., data format, file size, etc.)"""

    executor_type: str
    """Type of executor used (thread, process, or interpreter)"""

    qps: float
    """Queries per second (mean)"""

    ci_lower: float
    """Lower bound of 95% confidence interval for QPS"""

    ci_upper: float
    """Upper bound of 95% confidence interval for QPS"""

    date: str
    """When benchmark was run. ISO 8601 format."""

    cpu_percent: float
    """Average CPU utilization percentage during benchmark execution."""

    python_version: str = field(default=_PYTHON_VERSION)
    """Python version used for the benchmark"""

    free_threaded: bool = field(default=_FREE_THREADED)
    """Whether Python is running with free-threaded ABI."""


class ExecutorType(Enum):
    """ExecutorType()

    Supported executor types for concurrent execution."""

    THREAD = "thread"
    """Use :py:class:`~concurrent.futures.ThreadPoolExecutor`."""

    PROCESS = "process"
    """Use :py:class:`~concurrent.futures.ProcessPoolExecutor`."""

    INTERPRETER = "interpreter"
    """Use :py:class:`~concurrent.futures.InterpreterPoolExecutor`.

    Requires Python 3.14+.
    """


def _create_executor(executor_type: ExecutorType, max_workers: int) -> Executor:
    """Create an executor of the specified type.

    Args:
        executor_type: Type of executor to create
        max_workers: Maximum number of workers

    Returns:
        Executor instance

    Raises:
        ValueError: If ``executor_type`` is not supported
    """
    match executor_type:
        case ExecutorType.THREAD:
            return ThreadPoolExecutor(max_workers=max_workers)
        case ExecutorType.PROCESS:
            return ProcessPoolExecutor(max_workers=max_workers)
        case ExecutorType.INTERPRETER:
            from concurrent.futures import InterpreterPoolExecutor  # pyre-ignore[21]

            return InterpreterPoolExecutor(max_workers=max_workers)
        case _:
            raise ValueError(f"Unsupported executor type: {executor_type}")


def _verify_workers(executor: Executor, expected_workers: int) -> None:
    """Verify that the executor has created the expected number of workers.

    Args:
        executor: The executor to verify
        expected_workers: Expected number of workers

    Raises:
        RuntimeError: If the number of workers doesn't match expected
    """
    match executor:
        case ThreadPoolExecutor():
            actual_workers = len(executor._threads)
        case ProcessPoolExecutor():
            actual_workers = len(executor._processes)
        case _:
            raise ValueError(f"Unexpected executor type {type(executor)}")

    if actual_workers != expected_workers:
        raise RuntimeError(
            f"Expected {expected_workers} workers, but executor has {actual_workers}"
        )


def _warmup_executor(
    executor: Executor, func: Callable[[], T], num_iterations: int
) -> T:
    """Warmup the executor by running the function multiple times.

    Args:
        executor: The executor to warmup
        func: Function to run for warmup
        num_iterations: Number of warmup iterations

    Returns:
        Output from the last warmup iteration
    """
    futures = [executor.submit(func) for _ in range(num_iterations)]
    last_output: T | None = None
    for future in as_completed(futures):
        last_output = future.result()
    return last_output  # pyre-ignore[7]


class BenchmarkRunner:
    """Runner for executing benchmarks with configurable executors.

    This class provides a standardized way to run benchmarks with:

    - Warmup phase to exclude executor initialization overhead
    - Multiple runs for statistical confidence intervals
    - Support for different executor types

    The executor is initialized and warmed up in the constructor to exclude
    initialization overhead from benchmark measurements.

    Args:
        executor_type: Type of executor to use
            (``"thread"``, ``"process"``, or ``"interpreter"``)
        num_workers: Number of concurrent workers
        warmup_iterations: Number of warmup iterations (default: ``2 * num_workers``)
    """

    def __init__(
        self,
        executor_type: ExecutorType,
        num_workers: int,
        warmup_iterations: int | None = None,
    ) -> None:
        self._executor_type: ExecutorType = executor_type

        warmup_iters = (
            warmup_iterations if warmup_iterations is not None else 2 * num_workers
        )

        self._executor: Executor = _create_executor(executor_type, num_workers)

        _warmup_executor(self._executor, partial(time.sleep, 1), warmup_iters)
        _verify_workers(self._executor, num_workers)

    @property
    def executor_type(self) -> ExecutorType:
        """Get the executor type."""
        return self._executor_type

    def __enter__(self) -> "BenchmarkRunner":
        """Enter context manager."""
        return self

    def __exit__(self, exc_type: Any, exc_val: Any, exc_tb: Any) -> None:
        """Exit context manager and shutdown executor."""
        self._executor.shutdown(wait=True)

    def _run_iterations(
        self,
        func: Callable[[], T],
        iterations: int,
        num_runs: int,
    ) -> tuple[list[float], list[float], T]:
        """Run benchmark iterations and collect QPS and CPU utilization samples.

        Args:
            func: Function to benchmark (takes no arguments)
            iterations: Number of iterations per run
            num_runs: Number of benchmark runs

        Returns:
            Tuple of (list of QPS samples, list of CPU percent samples, last function output)
        """
        qps_samples: list[float] = []
        cpu_samples: list[float] = []
        last_output: T | None = None

        process = psutil.Process()

        for _ in range(num_runs):
            process.cpu_percent()
            t0 = time.perf_counter()
            futures = [self._executor.submit(func) for _ in range(iterations)]
            for future in as_completed(futures):
                last_output = future.result()
            elapsed = time.perf_counter() - t0
            cpu_percent = process.cpu_percent()
            qps_samples.append(iterations / elapsed)
            cpu_samples.append(cpu_percent / iterations)

        return qps_samples, cpu_samples, last_output  # pyre-ignore[7]

    def run(
        self,
        config: ConfigT,
        func: Callable[[], T],
        iterations: int,
        num_runs: int = 5,
        confidence_level: float = 0.95,
    ) -> tuple[BenchmarkResult[ConfigT], T]:
        """Run benchmark and return results with configuration.

        Args:
            config: Benchmark-specific configuration
            func: Function to benchmark (takes no arguments)
            iterations: Number of iterations per run
            num_runs: Number of benchmark runs for confidence interval calculation
                (default: ``5``)
            confidence_level: Confidence level for interval calculation (default: ``0.95``)

        Returns:
            Tuple of (``BenchmarkResult``, last output from function)
        """
        qps_samples, cpu_samples, last_output = self._run_iterations(
            func, iterations, num_runs
        )

        qps_mean = np.mean(qps_samples)
        qps_std = np.std(qps_samples, ddof=1)
        degrees_freedom = num_runs - 1
        confidence_interval = scipy.stats.t.interval(
            confidence_level,
            degrees_freedom,
            loc=qps_mean,
            scale=qps_std / np.sqrt(num_runs),
        )

        cpu_mean = np.mean(cpu_samples)

        date = datetime.now(timezone.utc).isoformat()

        result = BenchmarkResult(
            config=config,
            executor_type=self.executor_type.value,
            qps=float(qps_mean),
            ci_lower=float(confidence_interval[0]),
            ci_upper=float(confidence_interval[1]),
            date=date,
            cpu_percent=float(cpu_mean),
        )

        return result, last_output


def get_default_result_path(path: str, ext: str = ".csv") -> str:
    """Get the default result path with Python version appended."""
    base, _ = os.path.splitext(os.path.realpath(path))
    dirname = os.path.join(os.path.dirname(base), "data")
    filename = os.path.basename(base)
    version_suffix = (
        f"_{'.'.join(_PYTHON_VERSION.split('.')[:2])}{'t' if _FREE_THREADED else ''}"
    )
    return os.path.join(dirname, f"{filename}{version_suffix}{ext}")


def save_results_to_csv(
    results: list[BenchmarkResult[Any]],
    output_file: str,
) -> None:
    """Save benchmark results to a CSV file.

    Flattens the nested BenchmarkResult structure (config + performance metrics)
    into a flat CSV format. Each row contains both the benchmark configuration
    fields and the performance metrics.

    Args:
        results: List of BenchmarkResult instances
        output_file: Output file path for the CSV file
    """
    if not results:
        raise ValueError("No results to save")

    flattened_results = []
    for result in results:
        config_dict = asdict(result.config)
        # convert bool to int for slight readability improvement of raw CSV file
        config_dict = {
            k: (int(v) if isinstance(v, bool) else v) for k, v in config_dict.items()
        }
        flattened = {
            "date": result.date,
            "python_version": result.python_version,
            "free_threaded": int(result.free_threaded),
            **config_dict,
            "executor_type": result.executor_type,
            "qps": result.qps,
            "ci_lower": result.ci_lower,
            "ci_upper": result.ci_upper,
            "cpu_percent": result.cpu_percent,
        }
        flattened_results.append(flattened)

    # Get all field names from the first result
    fieldnames = list(flattened_results[0].keys())

    output_path = os.path.realpath(output_file)
    os.makedirs(os.path.dirname(output_file), exist_ok=True)
    with open(output_path, "w", newline="") as csvfile:
        # Write generated marker as first line
        # Note: Splitting the marker so as to avoid linter consider this file as generated file
        csvfile.write("# @")
        csvfile.write("generated\n")

        writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
        writer.writeheader()
        for result_dict in flattened_results:
            writer.writerow(result_dict)

    print(f"Results saved to {output_file}")


def load_results_from_csv(
    input_file: str,
    config_type: type[ConfigT],
) -> list[BenchmarkResult[ConfigT]]:
    """Load benchmark results from a CSV file.

    Reconstructs BenchmarkResult objects from the flattened CSV format created
    by :py:func:`save_results_to_csv`.
    Each row in the CSV is parsed into a :py:class:`BenchmarkResult`
    with the appropriate config type.

    Args:
        input_file: Input CSV file path
        config_type: The dataclass type to use for the config field

    Returns:
        List of BenchmarkResult instances with parsed config objects

    Raises:
        FileNotFoundError: If input_file does not exist
        ValueError: If CSV format is invalid or ``config_type`` is not a dataclass
    """
    if not hasattr(config_type, "__dataclass_fields__"):
        raise ValueError(f"config_type must be a dataclass, got {config_type}")
    fields: dict[str, Any] = config_type.__dataclass_fields__  # pyre-ignore[16]

    # Normalize input path and resolve symbolic links
    input_file = os.path.realpath(input_file)

    # Get the field names from the config dataclass
    config_fields = set(fields.keys())

    # Performance metric fields that are part of BenchmarkResult
    result_fields = {
        "executor_type",
        "qps",
        "ci_lower",
        "ci_upper",
        "date",
        "python_version",
        "free_threaded",
        "cpu_percent",
    }

    results: list[BenchmarkResult[ConfigT]] = []

    TRUES = ("true", "1", "yes")

    with open(input_file, newline="") as csvfile:
        reader = csv.DictReader((v for v in csvfile if not v.strip().startswith("#")))

        for row in reader:
            # Split row into config fields and result fields
            config_dict = {}
            result_dict = {}

            for key, value in row.items():
                if key in config_fields:
                    config_dict[key] = value
                elif key in result_fields:
                    result_dict[key] = value
                else:
                    # Unknown field - could be from config or result
                    # Try to infer based on whether it matches a config field name
                    config_dict[key] = value

            # Convert string values to appropriate types for config
            typed_config_dict = {}
            for field_name, field_info in fields.items():
                if field_name not in config_dict:
                    continue

                value = config_dict[field_name]
                field_type = field_info.type

                # Handle type conversions
                if field_type is int or field_type == "int":
                    typed_config_dict[field_name] = int(value)
                elif field_type is float or field_type == "float":
                    typed_config_dict[field_name] = float(value)
                elif field_type is bool or field_type == "bool":
                    typed_config_dict[field_name] = value.lower() in TRUES
                else:
                    # Keep as string or use the value as-is
                    typed_config_dict[field_name] = value

            result = BenchmarkResult(
                config=config_type(**typed_config_dict),
                executor_type=result_dict["executor_type"],
                qps=float(result_dict["qps"]),
                ci_lower=float(result_dict["ci_lower"]),
                ci_upper=float(result_dict["ci_upper"]),
                date=result_dict["date"],
                python_version=result_dict["python_version"],
                free_threaded=result_dict["free_threaded"].lower()
                in ("true", "1", "yes"),
                cpu_percent=float(result_dict.get("cpu_percent", 0.0)),
            )

            results.append(result)

    return results

API Reference

Functions

get_default_result_path(path: str, ext: str = '.csv') → str

Get the default result path with Python version appended.

load_results_from_csv(input_file: str, config_type: type[ConfigT]) → list[BenchmarkResult[ConfigT]]

Load benchmark results from a CSV file.

Reconstructs BenchmarkResult objects from the flattened CSV format created by save_results_to_csv(). Each row in the CSV is parsed into a BenchmarkResult with the appropriate config type.

Parameters:

• input_file – Input CSV file path

• config_type – The dataclass type to use for the config field

Returns:

List of BenchmarkResult instances with parsed config objects

Raises:

• FileNotFoundError – If input_file does not exist

• ValueError – If CSV format is invalid or config_type is not a dataclass

save_results_to_csv(results: list[BenchmarkResult[Any]], output_file: str) → None

Save benchmark results to a CSV file.

Flattens the nested BenchmarkResult structure (config + performance metrics) into a flat CSV format. Each row contains both the benchmark configuration fields and the performance metrics.

Parameters:

• results – List of BenchmarkResult instances

• output_file – Output file path for the CSV file

Classes

class BenchmarkRunner(executor_type: ExecutorType, num_workers: int, warmup_iterations: int | None = None)

Runner for executing benchmarks with configurable executors.

This class provides a standardized way to run benchmarks with:

• Warmup phase to exclude executor initialization overhead

• Multiple runs for statistical confidence intervals

• Support for different executor types

The executor is initialized and warmed up in the constructor to exclude initialization overhead from benchmark measurements.

Parameters:

• executor_type – Type of executor to use ("thread", "process", or "interpreter")

• num_workers – Number of concurrent workers

• warmup_iterations – Number of warmup iterations (default: 2 * num_workers)

property executor_type: ExecutorType

Get the executor type.

run(config: ConfigT, func: Callable[[], T], iterations: int, num_runs: int = 5, confidence_level: float = 0.95) → tuple[BenchmarkResult[ConfigT], T]

Run benchmark and return results with configuration.

Parameters:

• config – Benchmark-specific configuration

• func – Function to benchmark (takes no arguments)

• iterations – Number of iterations per run

• num_runs – Number of benchmark runs for confidence interval calculation (default: 5)

• confidence_level – Confidence level for interval calculation (default: 0.95)

Returns:

Tuple of (BenchmarkResult, last output from function)

class BenchmarkResult

Generic benchmark result containing configuration and performance metrics.

This class holds both the benchmark-specific configuration and the common performance statistics. It is parameterized by the config type, which allows each benchmark script to define its own configuration dataclass.

ci_lower: float

Lower bound of 95% confidence interval for QPS

ci_upper: float

Upper bound of 95% confidence interval for QPS

config: ConfigT

Benchmark-specific configuration (e.g., data format, file size, etc.)

cpu_percent: float

Average CPU utilization percentage during benchmark execution.

date: str

When benchmark was run. ISO 8601 format.

executor_type: str

Type of executor used (thread, process, or interpreter)

free_threaded: bool = False

Whether Python is running with free-threaded ABI.

python_version: str = '3.12.12'

Python version used for the benchmark

qps: float

Queries per second (mean)

class ExecutorType

Supported executor types for concurrent execution.

INTERPRETER = 'interpreter'

Use InterpreterPoolExecutor.

Requires Python 3.14+.

PROCESS = 'process'

Use ProcessPoolExecutor.

THREAD = 'thread'

Use ThreadPoolExecutor.