Differential Flame Graph enhancements: total-diff calculation and additional rendering customisation

[gielzak]

Overview:

This pull request introduces an opt-in total-diff calculation method for Differential Flame Graphs, complementing the existing self-diff approach. Additionally, it adds two new flags for diff rendering customisation.

TL;DR: see Demos section below

Proposed changes:

New flags:

• --totaldiff: aggregates diff values (i.e. uses total-diff not self-diff) when calculating differential hues
• --mindeltapc NUM: min delta (as % of function duration) to use diff hues (allows to reduce noise from insignificant changes)
• --fillopacity NUM: fill-opacity for drawn rectangles (SVG attribute; helps with method name readability)

Self-diff vs Total-diff:

Self-diff vs total-diff are based on self-time vs total-time. Basic example below:

• Base Trace (f1: 25%; f2: 25%; f3: 25%; f4: 25%):

  Benchmark.measure;PerfettoSdkBenchmark.basic;f0;f1 20
  Benchmark.measure;PerfettoSdkBenchmark.basic;f0;f1;getTime 5
  Benchmark.measure;PerfettoSdkBenchmark.basic;f0;f2 25
  Benchmark.measure;PerfettoSdkBenchmark.basic;f0;f3 25
  Benchmark.measure;PerfettoSdkBenchmark.basic;f0;f4 25

• Current Trace (f1: 20%; f2: 30%; f3: 10%; f4: 40%):

  Benchmark.measure;PerfettoSdkBenchmark.basic;f0;f1 20
  Benchmark.measure;PerfettoSdkBenchmark.basic;f0;f2 30
  Benchmark.measure;PerfettoSdkBenchmark.basic;f0;f3 10
  Benchmark.measure;PerfettoSdkBenchmark.basic;f0;f4 40

Self-diff doesn't highlight the performance gain in f1 (25% -> 20%), while total-diff captures this. See Demos section below for illustration.

Demos:

All hosted as interactive demos on https://gielzak.github.io/flamegraph-demos-totaldiff

• Basic example:

  • default (self-diff) vs total-diff

• Real code-change:

  • default (self-diff) vs total-diff
  • default (no-filter) vs min-delta 0.1%
  • default (fill-opacity 100%) vs fill-opacity 65%

[gielzak]

I cleaned-up the pull-request message a bit. It'd be great if you could review this and let me know what you think. I'd be happy to add tests (it didn't seem straightforward to extend existing ones) and/or adjust the changes to suit your coding conventions / vision for the tool.

The "total-diff" approach is more in line with what https://github.com/google/pprof would produce as diffs.

Note that in the demos (linked in the pull-request message) you can switch between graphs using the buttons at the top:

• base (flamegraph for the "diff base" trace)
• base-vs-current (differential flame graph with stacks from "base" trace and hues calculated based on the "current" trace)
• current (flamegraph for the "current" trace)
• current-vs-base (differential flame graph with stacks from "current" trace and hues calculated based on the "base" trace)

Also, note that the pull-request is split into small commits (hopefully easier to review).

Let me know what you think, and thank you for creating this awesome tool / approach to perf visualisation!