brintos

brintos / llvm-project-archived public Read only

0
0
Text · 5.1 KiB · f4936c0 Raw
168 lines · python
1"""2Static Analyzer qualification infrastructure.3 4This source file contains all the functionality related to benchmarking5the analyzer on a set projects.  Right now, this includes measuring6execution time and peak memory usage.  Benchmark runs analysis on every7project multiple times to get a better picture about the distribution8of measured values.9 10Additionally, this file includes a comparison routine for two benchmarking11results that plots the result together on one chart.12"""13 14import SATestUtils as utils15from SATestBuild import ProjectTester, stdout, TestInfo16from ProjectMap import ProjectInfo17 18import pandas as pd19from typing import List, Tuple20 21 22INDEX_COLUMN = "index"23 24 25def _save(data: pd.DataFrame, file_path: str):26    data.to_csv(file_path, index_label=INDEX_COLUMN)27 28 29def _load(file_path: str) -> pd.DataFrame:30    return pd.read_csv(file_path, index_col=INDEX_COLUMN)31 32 33class Benchmark:34    """35    Becnhmark class encapsulates one functionality: it runs the analysis36    multiple times for the given set of projects and stores results in the37    specified file.38    """39 40    def __init__(self, projects: List[ProjectInfo], iterations: int, output_path: str):41        self.projects = projects42        self.iterations = iterations43        self.out = output_path44 45    def run(self):46        results = [self._benchmark_project(project) for project in self.projects]47 48        data = pd.concat(results, ignore_index=True)49        _save(data, self.out)50 51    def _benchmark_project(self, project: ProjectInfo) -> pd.DataFrame:52        if not project.enabled:53            stdout(f" \n\n--- Skipping disabled project {project.name}\n")54            return55 56        stdout(f" \n\n--- Benchmarking project {project.name}\n")57 58        test_info = TestInfo(project)59        tester = ProjectTester(test_info, silent=True)60        project_dir = tester.get_project_dir()61        output_dir = tester.get_output_dir()62 63        raw_data = []64 65        for i in range(self.iterations):66            stdout(f"Iteration #{i + 1}")67            time, mem = tester.build(project_dir, output_dir)68            raw_data.append(69                {"time": time, "memory": mem, "iteration": i, "project": project.name}70            )71            stdout(72                f"time: {utils.time_to_str(time)}, "73                f"peak memory: {utils.memory_to_str(mem)}"74            )75 76        return pd.DataFrame(raw_data)77 78 79def compare(old_path: str, new_path: str, plot_file: str):80    """81    Compare two benchmarking results stored as .csv files82    and produce a plot in the specified file.83    """84    old = _load(old_path)85    new = _load(new_path)86 87    old_projects = set(old["project"])88    new_projects = set(new["project"])89    common_projects = old_projects & new_projects90 91    # Leave only rows for projects common to both dataframes.92    old = old[old["project"].isin(common_projects)]93    new = new[new["project"].isin(common_projects)]94 95    old, new = _normalize(old, new)96 97    # Seaborn prefers all the data to be in one dataframe.98    old["kind"] = "old"99    new["kind"] = "new"100    data = pd.concat([old, new], ignore_index=True)101 102    # TODO: compare data in old and new dataframes using statistical tests103    #       to check if they belong to the same distribution104    _plot(data, plot_file)105 106 107def _normalize(108    old: pd.DataFrame, new: pd.DataFrame109) -> Tuple[pd.DataFrame, pd.DataFrame]:110    # This creates a dataframe with all numerical data averaged.111    means = old.groupby("project").mean()112    return _normalize_impl(old, means), _normalize_impl(new, means)113 114 115def _normalize_impl(data: pd.DataFrame, means: pd.DataFrame):116    # Right now 'means' has one row corresponding to one project,117    # while 'data' has N rows for each project (one for each iteration).118    #119    # In order for us to work easier with this data, we duplicate120    # 'means' data to match the size of the 'data' dataframe.121    #122    # All the columns from 'data' will maintain their names, while123    # new columns coming from 'means' will have "_mean" suffix.124    joined_data = data.merge(means, on="project", suffixes=("", "_mean"))125    _normalize_key(joined_data, "time")126    _normalize_key(joined_data, "memory")127    return joined_data128 129 130def _normalize_key(data: pd.DataFrame, key: str):131    norm_key = _normalized_name(key)132    mean_key = f"{key}_mean"133    data[norm_key] = data[key] / data[mean_key]134 135 136def _normalized_name(name: str) -> str:137    return f"normalized {name}"138 139 140def _plot(data: pd.DataFrame, plot_file: str):141    import matplotlib142    import seaborn as sns143    from matplotlib import pyplot as plt144 145    sns.set_style("whitegrid")146    # We want to have time and memory charts one above the other.147    figure, (ax1, ax2) = plt.subplots(2, 1, figsize=(8, 6))148 149    def _subplot(key: str, ax: matplotlib.axes.Axes):150        sns.boxplot(151            x="project",152            y=_normalized_name(key),153            hue="kind",154            data=data,155            palette=sns.color_palette("BrBG", 2),156            ax=ax,157        )158 159    _subplot("time", ax1)160    # No need to have xlabels on both top and bottom charts.161    ax1.set_xlabel("")162 163    _subplot("memory", ax2)164    # The legend on the top chart is enough.165    ax2.get_legend().remove()166 167    figure.savefig(plot_file)168