Makie

PerfCheckerMakie renders saved measurements as Makie figures and interactive HTML. Rendering does not repeat the workload.

Install the renderer

PerfCheckerMakie 1.0.1 is registered in General and installs by name in a separate plotting environment:

import Pkg
Pkg.activate(joinpath(homedir(), ".julia", "environments", "perfchecker-plots"))
Pkg.add(["PerfChecker", "PerfCheckerMakie", "WGLMakie"])

The newer interactive features on this page use this site's renderer, PerfCheckerMakie 1.0.4, independently of Core 1.1.0. Version 1.0.4 is not registered in General; select this documentation's exact renderer source:

import Pkg
Pkg.activate(joinpath(homedir(), ".julia", "environments", "perfchecker-plots"))
Pkg.add(Pkg.PackageSpec(name = "PerfChecker", version = "1.1"))
Pkg.add(Pkg.PackageSpec(
    url = "https://github.com/Mirage-Interactive-Fr/PerfChecker.jl",
    subdir = "packages/PerfCheckerMakie", rev = "06b2b4d9b54c6c1c5c84ba346ff2317f896a20d9"))
Pkg.add("WGLMakie")

Published Core has a known ARM import limitation.

Open a saved bundle

Replace the path with a bundle printed by your completed suite run:

using PerfChecker, PerfCheckerMakie, WGLMakie

bundle = read_run_bundle("perf/results/ACTUAL-RUN/bundles/ACTUAL-BUNDLE")
catalog = plot_catalog(bundle)
foreach(entry -> println(entry["id"], " · ", entry["title"]), catalog)
isempty(catalog) && error("This bundle has no supported plots")
plot_id = first(catalog)["id"]
model = performance_plot(bundle, plot_id)
display(performance_figure(model))

The catalogue lists views supported by the saved observations: four-metric benchmark histories, distributions, allocation pies and attribution, CPU, wall-time and allocation flame graphs, or diagnostic series. A single-version bundle cannot supply a measured release history. Native TestItem JSON is a different format and is not accepted as a suite bundle.

Save interactive HTML

destination = "saved-performance.html"
ispath(destination) && error("Choose a new export filename")
write(destination, performance_plot_html(model))

Open the file in a browser. Point and frame selectors show exact saved values; zoom, pan and Fit work in the exported views. Normalized charts have metric toggles, version ranges and CSV/SVG downloads. Flame graphs retain frame inspection and nesting. Julia's live DataInspector callbacks are not exported.

This is native interactive output from the recorded DataStructures model:

Each curve has its own reference at 1. Read the statistic, reference, units and unavailable values in the exported plot before comparing curves.

Open the same interactive export · Download the serialized plot

Changing the view above reads the same saved values. It collects no measurements.

Normalized overlay

BenchmarkTools and Chairmarks each overlay elapsed time, GC, allocated bytes and allocation count on one ratio axis.

For each metric, take the minimum sample at each version and divide it by the smallest of those minima across the measured campaign. Each curve's minimum is 1; a ratio of 1.5 is 50% above that metric's minimum. The reference versions can differ. A metric that is zero throughout is shown at 1 by convention; missing observations remain unavailable. Inspect a point for its raw value and unit.

BenchmarkTools GC time is in seconds; Chairmarks GC is a fraction. Keep the collectors separate. Connecting segments guide the eye between recorded versions; they do not measure intermediate releases.

Models, bundles and static exports

A run bundle contains observations, definitions, provenance and integrity information. A plot model JSON is a presentation of selected observations; it does not replace the complete bundle or recover omitted measurements.

For static output, install CairoMakie and use its backend:

using CairoMakie
figure = performance_figure(model;
    figure_kwargs = (size = (1000, 600),),
    axis_kwargs = (backgroundcolor = :white,),
    plot_kwargs = (linewidth = 2,))
saveplot("comparison.svg", figure)
saveplot("comparison.png", figure; px_per_unit = 2)

Use fresh filenames. figure_kwargs, axis_kwargs and plot_kwargs target the figure, axes and primary recipe separately; attributes must suit the chosen recipe. SVG/PNG exports are static snapshots. Keep the bundle with shared HTML when the original evidence and provenance are needed.

Replay downloaded models and supplied examples

The shipped examples/kitchen-sink/replay.jl supplies saved_plot for a downloaded plot JSON. In that directory's prepared plotting environment:

using PerfChecker, PerfCheckerMakie, WGLMakie
include("replay.jl")
model = saved_plot("downloaded-plot.json")
write("replayed.html", performance_plot_html(model))

This re-renders existing values; it does not validate the original full bundle. For a complete saved run, julia --project=.controller/plots export.jl results/ACTUAL-RUN exports/history produces its catalogue, interactive HTML, plot JSON and terminal text. Prepare with julia setup.jl plots first.

Real allocation evidence is available as allocation-pie-5fa622ca80b2df9b in these three renderings:

Interactive HTML · Makie SVG · UnicodePlots text

View controls and shared assets

Allocation pies group sites below 5% of bytes into Other. Use performance_plot(bundle, pie_id; min_percentage = 2) to change the threshold, or 0 to disable grouping.

Flame width is recorded weight and depth is nesting, not chronology. Inspect the metric and unit, select narrow frames from the frame index, and use Fit or 0 to restore the full graph. Diagnostic series retain their recorded axes: payload size, memory observations and instruction counts are not releases.

For several HTML pages, pass both asset_directory and html_directory to performance_plot_html, then write its returned HTML. Keep the referenced assets and pages together. Omitting those options produces standalone output.