QtIndustrialWidgets 2.3.0
Modern, High-Performance Industrial Instruments & SCADA Widgets for Qt
Loading...
Searching...
No Matches
benchmarks

StripChart Performance Benchmarks

‍Note: Figures predate #11–#16 (rendering changes); regeneration tracked in #18.

‍AnalysisChart figures are in their own part at the end of this page: AnalysisChart Performance Benchmarks.

This page is the single source of performance figures for QtIndustrialWidgets::StripChart. The figures come from two consecutive runs of the same bench_StripChart build (benchmark source commit 75bb7d0) on the machine described below. Tables show the values of run 1 and, in the last column, the range of the medians of both runs. README, docs/strip_chart.md and the CHANGELOG quote this page or link to it.

All measurements render offscreen into a QPixmap with Qt's raster engine: they are CPU-only and do not include window-system composition or vsync.

Environment

Printed by the benchmark at start-up (identical in both runs):

Item Value
CPU 13th Gen Intel(R) Core(TM) i5-13600KF (20 logical CPUs, x86_64)
OS Ubuntu 24.04.1 LTS
Kernel linux 6.18.40.1-microsoft-standard-WSL2
Compiler GCC 13.3.0
Build NDEBUG defined, OPTIMIZE defined
Qt 6.4.2 (compiled against 6.4.2)
QPA platform offscreen
Render target QPixmap via QWidget::render() (CPU raster)
Device pixel ratio 1
Widget size 1974 × 1080 px
Plot width W_dev 1920 device px
Library version 2.1.1
Benchmark source commit 75bb7d0ab3025439dda8e7ec1978837d3c76f0c4
Heap allocation counting malloc/calloc/realloc interposed (glibc)

Not printed by the benchmark: the compiler flags come from the CMake Release configuration (-O3 -DNDEBUG -std=c++17 -Wall -Wextra -Wpedantic -fPIC). The machine runs Ubuntu under WSL2 on a Windows host; the CPU is a hybrid design (performance and efficiency cores).

Commands

cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release \
-DBUILD_TESTS=ON -DBUILD_EXAMPLES=ON -DBUILD_DOCS=OFF \
-DBUILD_DESIGNER_PLUGIN=OFF -DBUILD_BENCHMARKS=ON
cmake --build build
for k in 1 2; do
QT_QPA_PLATFORM=offscreen ./build/bin/bench_StripChart -o bench_log$k.txt,txt > bench_run$k.md
done
python3 benchmarks/merge_runs.py bench_run1.md bench_run2.md > bench_results.md

bench_results.md contains the environment and results tables below verbatim; progress lines go to stderr. merge_runs.py stops if the two runs did not draw the same number of trace pixels.

Method

  • Widget: 1974 × 1080 px at DPR 1, plot width W_dev = 1920 device px, DecimationMode::Auto, default channel pen width 1.8 px, autoscale off unless stated. Time-mode cases use a 10 s window filled with N uniformly spaced samples per channel; the ring-buffer capacity is N. With N ≥ 10k every case draws decimated traces.
  • Signals are functions of time t = i · dt, so every N draws the same waveform in the window: SmoothSine (sine), BroadbandNoise (the same sine plus per-sample noise and an alternating ±3 offset, so every pixel column has a min/max envelope), SpikesAndNaN (the sine with a one-sample spike every 100 ms and a 50 µs NaN dropout every 250 ms; at 10k samples a dropout covers at most one sample). At N = 1M the samples are bit-identical to the index-based generator used before.
  • Trace px: trace pixels in the last rendered frame, counted as pixels inside the plot rectangle whose colour equals a channel colour (decimated traces are drawn without antialiasing). It measures how much the rasteriser has to draw, so growth in time can be compared with growth in drawn pixels.
  • Paint at steady state: data inserted once, then 5 warm-up renders and 30 measured renders; min, median and p95 (nearest rank) of the 30 render times.
  • Paint with streaming: before every render, 1/60 s of acquisition at the case's sample rate (N / 10 s) is appended to every channel with addUniformSamples(), so the ring buffer evicts as much as it receives. The appended blocks continue the same signal (verified: after 35 frames the streamed image is pixel-identical to a chart that received the same samples in one block). 5 warm-up and 30 measured frames; render and append time are reported separately.
  • Heap allocations: counted in the streaming case (run 1) by interposing glibc malloc / calloc / realloc in the benchmark executable. The count is process-wide: it includes QPainter, text and pixmap allocations, not only the library's own.
  • First paint after resize: 10 independent trials; each builds a fresh chart, renders at 1000 × 800, resizes to 1974 × 1080 and times the next render (cold path: decimation streams are rebuilt).
  • Insertion: real call counts, never extrapolated; 10 trials each. The addSample cases run against a ring buffer that already holds 1M samples (capacity 1M, so every call evicts one sample).

Run-to-run variation (bimodal timings)

On this machine render times are bimodal: the same case runs either in a slow mode or in a mode roughly twice as fast, within one run (compare min and median) and between runs (last column). The draw workload is the same in both modes: trace pixel counts are identical between runs, and steady and streaming frames of the same case draw the same pixels. In fast runs the append time, which involves no rendering, drops as well, which points to the whole thread running faster (for example scheduling on a different core type of the hybrid CPU under WSL2, or a different clock frequency). This explanation has not been verified. Small differences between cases on this page are therefore not meaningful, and verdicts below are given for both runs.

Results

Paint at steady state (5 warm-up + 30 measured renders, ms)

Channels Samples / ch Signal X axis Trace px Min Median p95 Median, 2 runs (min–max)
1 10k SmoothSine Time 323992 1.74 1.89 2.18 1.89–1.95
1 10k SmoothSine SampleIndex 325795 1.63 1.84 1.93 1.84–1.88
1 10k BroadbandNoise Time 301104 1.94 2.06 2.35 2.06–2.08
1 10k BroadbandNoise SampleIndex 292801 1.88 1.97 2.24 1.97–2.02
1 10k SpikesAndNaN Time 373488 1.89 2.01 2.18 2.01–2.08
1 10k SpikesAndNaN SampleIndex 395533 1.94 2.07 2.32 2.02–2.07
1 100k SmoothSine Time 324093 1.73 1.91 2.20 1.85–1.91
1 100k SmoothSine SampleIndex 325481 1.68 1.86 2.00 1.84–1.86
1 100k BroadbandNoise Time 437060 2.51 2.64 2.94 2.64–2.66
1 100k BroadbandNoise SampleIndex 360977 2.50 2.58 2.72 2.58–2.65
1 100k SpikesAndNaN Time 375778 1.89 2.04 2.26 2.04–2.06
1 100k SpikesAndNaN SampleIndex 397424 1.95 2.10 2.53 2.03–2.10
1 1M SmoothSine Time 324031 1.70 1.90 2.18 1.90–1.90
1 1M SmoothSine SampleIndex 325404 1.69 1.91 2.14 1.91–1.93
1 1M BroadbandNoise Time 487658 2.80 2.96 3.15 2.96–3.99
1 1M BroadbandNoise SampleIndex 371916 2.75 2.89 5.47 2.89–3.39
1 1M SpikesAndNaN Time 367496 1.92 2.08 2.49 2.08–2.11
1 1M SpikesAndNaN SampleIndex 398302 1.94 2.12 2.37 2.11–2.12
1 4M SmoothSine Time 323999 1.74 1.95 2.54 1.95–1.98
1 4M SmoothSine SampleIndex 325391 1.78 1.92 2.24 1.84–1.92
1 4M BroadbandNoise Time 498762 2.87 3.08 4.09 3.08–3.99
1 4M BroadbandNoise SampleIndex 378141 2.85 3.88 6.64 3.04–3.88
1 4M SpikesAndNaN Time 367562 1.89 2.27 3.45 2.14–2.27
1 4M SpikesAndNaN SampleIndex 398290 1.94 2.18 3.11 2.12–2.18
4 10k SmoothSine Time 1091720 4.94 5.31 6.54 5.31–5.33
4 10k SmoothSine SampleIndex 1100099 4.82 5.33 7.30 5.33–5.39
4 10k BroadbandNoise Time 991391 5.96 6.40 8.59 6.25–6.40
4 10k BroadbandNoise SampleIndex 976091 5.60 6.03 7.57 5.77–6.03
4 10k SpikesAndNaN Time 1138214 5.89 6.28 7.92 6.25–6.28
4 10k SpikesAndNaN SampleIndex 1176758 5.85 6.22 7.77 6.22–9.24
4 100k SmoothSine Time 1091804 5.13 5.53 6.89 5.53–6.15
4 100k SmoothSine SampleIndex 1099809 5.12 5.51 8.07 5.27–5.51
4 100k BroadbandNoise Time 1284392 8.17 11.92 19.89 8.52–11.92
4 100k BroadbandNoise SampleIndex 1135608 7.83 10.88 20.40 8.58–10.88
4 100k SpikesAndNaN Time 1143538 6.11 10.78 12.86 6.13–10.78
4 100k SpikesAndNaN SampleIndex 1181055 9.59 11.47 12.69 5.98–11.47
4 1M SmoothSine Time 1091747 5.16 9.19 10.25 5.15–9.19
4 1M SmoothSine SampleIndex 1099842 4.91 5.17 6.91 5.17–5.27
4 1M BroadbandNoise Time 1365399 16.39 23.08 25.10 11.42–23.08
4 1M BroadbandNoise SampleIndex 1154653 10.73 23.10 25.30 16.01–23.10
4 1M SpikesAndNaN Time 1141303 7.91 10.99 12.16 6.37–10.99
4 1M SpikesAndNaN SampleIndex 1181836 5.65 6.01 6.35 6.01–10.42
4 4M SmoothSine Time 1091775 5.58 9.48 10.48 8.95–9.48
4 4M SmoothSine SampleIndex 1099894 5.18 8.29 10.04 5.88–8.29
4 4M BroadbandNoise Time 1380651 16.95 22.75 24.21 22.75–23.74
4 4M BroadbandNoise SampleIndex 1165839 11.84 23.99 25.10 23.99–24.05
4 4M SpikesAndNaN Time 1144058 8.88 10.09 12.11 9.80–10.09
4 4M SpikesAndNaN SampleIndex 1181901 5.96 7.95 12.45 7.95–10.38

Paint with streaming (1/60 s of data appended to every channel before each render; 5 warm-up + 30 measured frames, ms)

Channels Samples / ch Signal X axis New samples / ch / frame Trace px Render min Render median Render p95 Append median (all ch) Median, 2 runs (min–max)
1 10k SmoothSine Time 17 323785 1.75 1.90 2.10 0.00 1.87–1.90
1 10k SmoothSine SampleIndex 17 325687 1.68 1.86 1.97 0.00 1.79–1.86
1 10k BroadbandNoise Time 17 301138 2.00 2.12 2.34 0.00 2.05–2.12
1 10k BroadbandNoise SampleIndex 17 292659 1.87 2.04 2.16 0.00 1.95–2.04
1 10k SpikesAndNaN Time 17 372658 1.95 2.10 2.64 0.00 2.09–2.10
1 10k SpikesAndNaN SampleIndex 17 393724 1.96 2.11 2.72 0.00 2.10–2.11
1 100k SmoothSine Time 167 323882 1.69 1.92 2.22 0.00 1.87–1.92
1 100k SmoothSine SampleIndex 167 325297 1.59 1.86 2.09 0.00 1.86–1.89
1 100k BroadbandNoise Time 167 436785 2.50 2.64 2.89 0.00 2.64–2.73
1 100k BroadbandNoise SampleIndex 167 359342 2.43 2.58 3.05 0.00 2.58–2.71
1 100k SpikesAndNaN Time 167 374845 1.85 2.01 2.15 0.00 2.01–2.08
1 100k SpikesAndNaN SampleIndex 167 394614 1.85 2.00 2.10 0.00 2.00–2.09
1 1M SmoothSine Time 1667 323765 1.74 1.91 2.30 0.01 1.91–1.94
1 1M SmoothSine SampleIndex 1667 325134 1.82 2.13 4.76 0.01 2.07–2.13
1 1M BroadbandNoise Time 1667 487478 2.83 3.00 3.84 0.01 2.97–3.00
1 1M BroadbandNoise SampleIndex 1667 368080 2.85 3.27 6.01 0.01 3.06–3.27
1 1M SpikesAndNaN Time 1667 366869 2.02 2.11 2.29 0.01 2.08–2.11
1 1M SpikesAndNaN SampleIndex 1667 398536 2.07 2.87 10.28 0.01 2.34–2.87
1 4M SmoothSine Time 6667 323731 1.70 1.90 3.11 0.03 1.82–1.90
1 4M SmoothSine SampleIndex 6667 325055 2.02 4.06 39.41 0.12 3.45–4.06
1 4M BroadbandNoise Time 6667 498604 2.90 3.88 5.33 0.10 3.80–3.88
1 4M BroadbandNoise SampleIndex 6667 369437 3.73 7.36 43.36 0.12 7.20–7.36
1 4M SpikesAndNaN Time 6667 366934 1.86 2.08 2.81 0.03 2.05–2.08
1 4M SpikesAndNaN SampleIndex 6667 398982 2.10 3.71 38.04 0.07 3.68–3.71
4 10k SmoothSine Time 17 1091820 5.20 8.41 10.65 0.00 8.41–8.53
4 10k SmoothSine SampleIndex 17 1100637 5.09 5.57 8.89 0.00 5.57–8.26
4 10k BroadbandNoise Time 17 990855 6.04 6.42 8.87 0.00 6.42–12.11
4 10k BroadbandNoise SampleIndex 17 977157 5.59 5.97 7.08 0.00 5.97–10.92
4 10k SpikesAndNaN Time 17 1138621 5.79 6.36 8.57 0.00 6.36–9.81
4 10k SpikesAndNaN SampleIndex 17 1181589 5.79 6.46 7.76 0.00 6.45–6.46
4 100k SmoothSine Time 167 1091847 5.33 8.35 9.37 0.01 5.69–8.35
4 100k SmoothSine SampleIndex 167 1100394 5.02 8.37 11.10 0.01 5.42–8.37
4 100k BroadbandNoise Time 167 1284222 8.47 9.93 12.84 0.01 9.93–20.83
4 100k BroadbandNoise SampleIndex 167 1132627 8.41 20.52 22.10 0.02 20.52–20.92
4 100k SpikesAndNaN Time 167 1143865 9.36 10.29 12.06 0.01 10.29–10.43
4 100k SpikesAndNaN SampleIndex 167 1179940 6.46 10.23 13.12 0.01 10.23–11.90
4 1M SmoothSine Time 1667 1091652 7.62 8.21 10.05 0.12 8.02–8.21
4 1M SmoothSine SampleIndex 1667 1100227 5.41 10.27 49.18 0.12 9.77–10.27
4 1M BroadbandNoise Time 1667 1365550 16.44 22.84 24.03 0.12 22.84–23.31
4 1M BroadbandNoise SampleIndex 1667 1146339 21.03 24.55 62.95 0.12 24.55–24.82
4 1M SpikesAndNaN Time 1667 1140473 5.79 10.11 12.34 0.12 6.22–10.11
4 1M SpikesAndNaN SampleIndex 1667 1183007 9.19 12.81 51.77 0.12 12.78–12.81
4 4M SmoothSine Time 6667 1091699 5.08 8.30 9.91 0.48 5.27–8.30
4 4M SmoothSine SampleIndex 6667 1100083 8.40 10.78 157.64 0.48 10.72–10.78
4 4M BroadbandNoise Time 6667 1380895 20.94 23.84 24.84 0.49 23.66–23.84
4 4M BroadbandNoise SampleIndex 6667 1149178 21.22 24.87 171.61 0.49 24.87–24.88
4 4M SpikesAndNaN Time 6667 1143351 5.59 6.56 12.07 0.25 6.56–9.86
4 4M SpikesAndNaN SampleIndex 6667 1183441 10.93 12.66 161.08 0.49 12.66–12.69

Heap allocations in the streaming case (per measured frame; median / max over 30 frames)

Channels Samples / ch Signal X axis Allocations per render KiB allocated per render Allocations per append (all ch)
1 10k SmoothSine Time 127 / 131 141.1 / 150.1 0 / 0
1 10k SmoothSine SampleIndex 125 / 130 138.1 / 148.1 0 / 0
1 10k BroadbandNoise Time 126 / 131 266.1 / 276.1 0 / 0
1 10k BroadbandNoise SampleIndex 127 / 130 268.1 / 274.1 0 / 0
1 10k SpikesAndNaN Time 165 / 169 686.1 / 694.2 0 / 0
1 10k SpikesAndNaN SampleIndex 165 / 169 686.1 / 694.2 0 / 0
1 100k SmoothSine Time 125 / 132 263.0 / 278.2 0 / 0
1 100k SmoothSine SampleIndex 125 / 131 138.1 / 150.2 0 / 0
1 100k BroadbandNoise Time 125 / 132 264.0 / 278.2 0 / 0
1 100k BroadbandNoise SampleIndex 126 / 132 265.0 / 278.2 0 / 0
1 100k SpikesAndNaN Time 164 / 170 683.1 / 696.2 0 / 0
1 100k SpikesAndNaN SampleIndex 164 / 168 684.1 / 692.1 0 / 0
1 1M SmoothSine Time 125 / 132 263.0 / 278.2 0 / 0
1 1M SmoothSine SampleIndex 126 / 132 266.1 / 278.2 0 / 0
1 1M BroadbandNoise Time 125 / 132 264.0 / 278.2 0 / 0
1 1M BroadbandNoise SampleIndex 126 / 132 265.0 / 278.2 0 / 0
1 1M SpikesAndNaN Time 202 / 211 1320.0 / 1332.1 0 / 0
1 1M SpikesAndNaN SampleIndex 202 / 209 1319.0 / 1334.2 0 / 0
1 4M SmoothSine Time 126 / 132 266.1 / 278.2 0 / 0
1 4M SmoothSine SampleIndex 123 / 130 260.0 / 274.1 0 / 0
1 4M BroadbandNoise Time 125 / 132 264.0 / 278.2 0 / 0
1 4M BroadbandNoise SampleIndex 123 / 129 259.0 / 272.1 0 / 0
1 4M SpikesAndNaN Time 203 / 209 1320.0 / 1334.2 0 / 0
1 4M SpikesAndNaN SampleIndex 200 / 207 1316.0 / 1330.1 0 / 0
4 10k SmoothSine Time 210 / 216 482.8 / 494.9 0 / 0
4 10k SmoothSine SampleIndex 211 / 218 484.8 / 498.9 0 / 0
4 10k BroadbandNoise Time 215 / 222 996.8 / 1010.9 0 / 0
4 10k BroadbandNoise SampleIndex 214 / 221 994.8 / 1008.9 0 / 0
4 10k SpikesAndNaN Time 363 / 369 2659.8 / 2672.9 0 / 0
4 10k SpikesAndNaN SampleIndex 362 / 368 2658.8 / 2670.9 0 / 0
4 100k SmoothSine Time 215 / 222 995.8 / 1010.9 0 / 0
4 100k SmoothSine SampleIndex 209 / 216 480.8 / 494.9 0 / 0
4 100k BroadbandNoise Time 214 / 220 994.8 / 1006.9 0 / 0
4 100k BroadbandNoise SampleIndex 213 / 217 991.8 / 1000.8 0 / 0
4 100k SpikesAndNaN Time 361 / 367 2655.8 / 2668.9 0 / 0
4 100k SpikesAndNaN SampleIndex 362 / 368 2658.8 / 2670.9 0 / 0
4 1M SmoothSine Time 213 / 220 992.8 / 1006.9 0 / 0
4 1M SmoothSine SampleIndex 212 / 218 990.8 / 1002.9 0 / 0
4 1M BroadbandNoise Time 213 / 219 992.8 / 1004.9 0 / 0
4 1M BroadbandNoise SampleIndex 213 / 218 992.8 / 1002.9 0 / 0
4 1M SpikesAndNaN Time 521 / 528 5216.8 / 5230.9 0 / 0
4 1M SpikesAndNaN SampleIndex 520 / 525 5214.8 / 5224.8 0 / 0
4 4M SmoothSine Time 213 / 221 992.8 / 1008.9 0 / 0
4 4M SmoothSine SampleIndex 212 / 220 990.8 / 1006.9 0 / 0
4 4M BroadbandNoise Time 212 / 219 990.8 / 1004.9 0 / 0
4 4M BroadbandNoise SampleIndex 212 / 218 989.8 / 1002.9 0 / 0
4 4M SpikesAndNaN Time 521 / 528 5216.8 / 5230.9 0 / 0
4 4M SpikesAndNaN SampleIndex 520 / 528 5214.8 / 5230.9 0 / 0

First paint after resize (cold path)

Case Work per trial Trials Min (ms) Median (ms) p95 (ms) Median per call (µs) Median, 2 runs (min–max)
4 ch × 1M BroadbandNoise, Time; render at 1000×800, resize to 1974×1080, time the next render 1 render 10 42.05 66.64 70.86 – 64.78–66.64

Insertion

Case Work per trial Trials Min (ms) Median (ms) p95 (ms) Median per call (µs) Median, 2 runs (min–max)
addSample, autoscale ON, 1M samples in buffer 100000 calls 10 2.95 3.23 3.43 0.0323 3.22–3.23
addSample, autoscale OFF, 1M samples in buffer 100000 calls 10 3.20 3.32 3.65 0.0332 3.23–3.32
addUniformSamples, one 1M-sample block into an empty channel 1 call (1M samples) 10 8.85 9.03 9.26 – 6.10–9.03
addDataPoint (legacy), SampleIndex, autoscale OFF, empty channel 100000 calls 10 3.22 3.25 3.87 0.0325 3.24–3.25

Targets

The three acceptance targets set for 2.1.0, evaluated on the steady-state median as originally defined, for both runs. A target is marked PASS only if it is met in both runs. The targets are unchanged.

Target Run 1 Run 2 Result
4 ch × 1M noisy ≤ 8 ms per frame Time 23.08 ms; SampleIndex 23.10 ms Time 11.42 ms; SampleIndex 16.01 ms FAIL (both runs)
Paint-time growth ≤ 25 % from 100k to 4M samples (1 ch, noisy) Time 2.64 → 3.08 ms = +16.7 %; SampleIndex 2.58 → 3.88 ms = +50.4 % Time 2.66 → 3.99 ms = +50.0 %; SampleIndex 2.65 → 3.04 ms = +14.7 % FAIL (not met in every run)
addSample with autoscale ON ≤ 1 µs per call 0.0323 µs 0.0322 µs PASS (both runs)

Streaming values for the first target (render median): run 1 Time 22.84 ms, SampleIndex 24.55 ms; run 2 Time 23.31 ms, SampleIndex 24.82 ms. addSample figures are medians of 10 trials × 100 000 calls with 1M samples in the buffer.

Growth target. Over the same range the number of trace pixels grows by +14.1 % in Time mode (437060 → 498762) and +4.8 % in SampleIndex mode (360977 → 378141), identical in both runs. In run 1, Time mode, the growth in time (+17 %) follows the growth in trace pixels (+14 %), so there is no evidence of a library cost that depends on N beyond the image itself. The two failing values (run 1 SampleIndex, run 2 Time) come from a 4M median that fell into the slow mode of the machine (3.88 and 3.99 ms, against 3.04 and 3.08 ms for the same cases in the other run). Each axis passes in one run and fails in the other, so the target is marked FAIL; a final verdict needs a measurement that is insensitive to the bimodal timings.

Dominant cost of the 4 ch × 1M case. A callgrind profile (Time, steady state, collection limited to StripChart::paintEvent; data identical to this page at N = 1M) attributes 93.8 % of the instructions to QPainter::drawPolyline → QRasterPaintEngine::drawPolygon, i.e. rasterising the decimated trace polylines inside Qt; extracting buckets from the decimation stream is below 1 %. The hot function inside Qt has no symbol in the Qt build used here. Profile details: #5 (comment). Follow-up: #5, #6.

Other findings, tracked separately:

  • Streaming in SampleIndex mode shows periodic slow frames (p95 column of the streaming table): #7.
  • With NaN gaps, allocations per frame grow with the number of visible gaps (SpikesAndNaN rows of the allocation table): #8.

Baseline: 2.0.1, re-measured in the same session

Measured in the same session, on the same machine, with the benchmark at commit e5595b2 compiled against the library at tag v2.0.1 (commit f2fdcb7), same compiler and Release flags. That benchmark version still used the previous, index-based signal generator; at N = 1M, the only sample count used below, its samples are bit-identical to those of the current generator, so the two columns draw the same data. The 2.1.1 column is run 1 of this page, with the range of the two runs in brackets.

2.0.1 has no benchmarks/ directory, so the source file was copied into a v2.0.1 worktree and built with a small CMake project that adds the worktree with add_subdirectory() and links QtIndustrialWidgets, Qt6::Widgets and Qt6::Test. The benchmark reports that build as f2fdcb7 (bench_StripChart.cpp differs from this commit), library version 2.0.1.

git worktree add --detach ../wt-v2.0.1 v2.0.1
git show e5595b2:benchmarks/bench_StripChart.cpp > ../wt-v2.0.1/benchmarks/bench_StripChart.cpp
# configure the small CMake project with -G Ninja -DCMAKE_BUILD_TYPE=Release and build it, then:
BENCH_ADDSAMPLE_CALLS=16000 BENCH_INSERT_TRIALS=1 QT_QPA_PLATFORM=offscreen ./bench_StripChart \
-o base_log.txt,txt \
paintSteady:1ch_1M_BroadbandNoise_Time paintSteady:4ch_1M_BroadbandNoise_Time \
insertionThroughput_addSample_autoscaleOn > base_results.md

The addSample case uses 16000 calls in 1 trial on 2.0.1 (about one minute) instead of 10 × 100 000, because each call costs milliseconds on 2.0.1.

Case 2.0.1: min / median / p95 2.1.1: min / median / p95 (median range, 2 runs)
Paint, steady, 1 ch × 1M BroadbandNoise, Time 5855.47 / 7413.47 / 7949.81 ms 2.80 / 2.96 / 3.15 ms (2.96–3.99)
Paint, steady, 4 ch × 1M BroadbandNoise, Time 23221.93 / 31832.63 / 32707.65 ms 16.39 / 23.08 / 25.10 ms (11.42–23.08)
addSample, autoscale ON, 1M samples in buffer (median per call) 3635.9857 µs (16000 calls, 1 trial) 0.0323 µs (100000 calls, 10 trials; run 2: 0.0322 µs)

Earlier 2.0.1 baseline figures in this file and in the CHANGELOG, including an extrapolated addSample figure, were not reproducible and have been removed. For an earlier 2.0.0 figure that came from a lighter case, see the note on issue #4.

AnalysisChart Performance Benchmarks

The single source of performance figures for QtIndustrialWidgets::AnalysisChart (#25). The figures come from one run of bench_AnalysisChart (benchmark source commit 4232d72) on the machine of the StripChart part above, with docs/analysis_chart.md, README and CHANGELOG quoting this part.

Target

#25: a redraw after a zoom or pan step takes ≤ 16.7 ms for 4 channels × 10 M samples at a 1920 device px plot. A case passes when the p95 of its 30 measured redraws is ≤ 16.7 ms.

Result: not met. 6 of the 7 cases FAIL; only panning at 1/100 of the record (100 k samples per channel in the view) passes. The cause is below.

Command

cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release \
-DBUILD_TESTS=ON -DBUILD_EXAMPLES=ON -DBUILD_DOCS=OFF \
-DBUILD_DESIGNER_PLUGIN=OFF -DBUILD_BENCHMARKS=ON
cmake --build build --clean-first
QT_QPA_PLATFORM=offscreen ./build/bin/bench_AnalysisChart > bench_analysis.md

bench_analysis.md contains the tables below verbatim; progress lines go to stderr.

Method

  • Widget: AnalysisChart at DPR 1, resized so that the plot is exactly 1920 device px wide (1974 × 1080 px), DecimationMode::Auto, readout band hidden, rendered into a QPixmap with QWidget::render() (CPU raster, offscreen, no composition or vsync).
  • Data: 4 channels × 10 M samples every 10 µs (100 s). The signals are those of bench_StripChart, sample for sample (its reference interval is 10 µs): SpikesAndNaN for all zoom and pan cases, SmoothSine and BroadbandNoise for one more full-view case each. Over 100 s the 62.8 ms sine of these signals fills every pixel column of the full view with a full-height envelope.
  • Cases: 5 warm-up renders, then 30 measured redraws; min, median and p95 (nearest rank). A measured redraw of a zoom or pan case is one input step delivered to the widget synchronously, then one render, timed together:
    • Zoom in / Zoom out: a wheel notch (QWheelEvent, ×1.25) at the centre of the plot, 30 steps from the full view in, then 30 steps back out;
    • Pan: arrow keys (Qt::Key_Right 15 times, then Qt::Key_Left 15 times, 10 % of the view each) at a view of 1/10 and of 1/100 of the record.
  • setChannelData(): wall time of each of the 4 calls; memory as the growth of the resident set size of the process across the 4 calls (the benchmark's own buffers freed), and the bytes of the pyramid.
  • The timings of this machine are bimodal between runs (see the StripChart part); the spread between min and p95 within one case shows it.

Results

Environment

Item Value
CPU 13th Gen Intel(R) Core(TM) i5-13600KF (20 logical CPUs, x86_64)
OS Ubuntu 24.04.1 LTS
Kernel linux 6.18.54.1-microsoft-standard-WSL2
Compiler GCC 13.3.0
Build NDEBUG defined, OPTIMIZE defined
Qt 6.4.2 (compiled against 6.4.2)
QPA platform offscreen
Render target QPixmap via QWidget::render() (CPU raster)
Device pixel ratio 1
Widget size 1974 × 1080 px
Plot width W_dev 1920 device px
Data 4 channels × 10000000 samples, every 10 µs
Library version 2.2.0
Benchmark source commit 4232d72e2a318594236d3c862506302a0fe65996

Redraw after a zoom or pan step (target ≤ 16.7 ms at p95)

Case Samples per channel in view Decimated channels min (ms) median (ms) p95 (ms) Target
Full view, redraw only (SpikesAndNaN) 1e+07 4 / 4 33.46 42.07 65.22 FAIL
Zoom in, wheel x1.25 per step 1e+07 → 1.238e+04 4 / 4 3.70 11.90 35.37 FAIL
Zoom out, wheel x1.25 per step 1.238e+04 → 1e+07 4 / 4 3.99 8.99 38.40 FAIL
Pan, arrow keys 10 % per step, 1/10 of the record 1e+06 4 / 4 12.90 19.10 20.21 FAIL
Pan, arrow keys 10 % per step, 1/100 of the record 1e+05 4 / 4 7.06 10.06 10.81 PASS
Full view, redraw only (SmoothSine) 1e+07 4 / 4 45.53 62.88 91.55 FAIL
Full view, redraw only (BroadbandNoise) 1e+07 4 / 4 40.45 84.56 104.60 FAIL

setChannelData()

Item Value
Time per call (4 calls of 10000000 samples) min 121.85 ms, median 149.02 ms
Stored samples (8 B timestamp + 8 B value) 610.4 MiB
Pyramid 81.1 MiB (2.13 B/sample, 13.3 % of the samples)
Resident set growth across the 4 calls 691.9 MiB

Cause of the miss (profile)

Measured during development with temporary timers around the phases of paintEvent() (not committed), on the same machine and data, full view, SpikesAndNaN, per frame:

Phase Time per frame
Buckets from the pyramid (4 channels × 1920 buckets) 7.5–10.2 ms
Segments from the buckets 0.2–0.4 ms
Drawing the decimated polylines 17–30 ms
Whole frame 27–43 ms
  • Drawing dominates. Each M4 column is drawn first → min → max → last, so its vertical extent is stroked twice (up and back down); with an envelope in every column that is several million pixels written per frame for 4 channels, of which about 680 k remain visible. A separate QPainter test with the same pen showed the cost does not depend on the number of drawPolyline() calls (the same polylines in 4 or in 1536 calls: 7.3 ms each), so it follows the ink. StripChart draws its decimated traces the same way; drawing fewer pixels (e.g. one vertical span per column) would change the rendered pixels of both charts and is left to a decision of the maintainer.
  • Buckets. 4 × 1920 buckets each read up to two partial blocks of raw samples and about 2 log2(blocks) pyramid nodes; at 10 M samples these reads are mostly cache misses. A compact index of the first timestamp of every block replaced the binary search over the whole timestamp array (searches 4.4 → 3.7 ms per frame in a micro-benchmark).
  • The zoom cases fail at p95 because their first steps are close to the full view; panning at 1/10 of the record (1 M samples per channel in view) fails with a p95 of 20.21 ms.