D Donkeh docs
GitHub

Using the performance tab

The Performance tab turns the last few seconds of frame work into a small dashboard. It is designed to show where a frame goes, not to replace a profiler.

What is sampled

The host samples while the Performance tab is visible, at most every 0.25 seconds. Each graph keeps up to 120 samples. The panel refreshes roughly once per second, so the numbers are recent but not instantaneous.

Sampling pauses when hidden Leave the tab open while reproducing a slow interaction. Switching away stops the sample stream, so the graph will not contain work performed while another tab was active.

The summary cards

CardMeaning
FPSWhole frames per second derived from the latest wall-clock frame time. The small line also shows the frame time in milliseconds.
CPUProcess-wide CPU usage percentage and the current thread count. The first sample has no baseline, so it reports zero.
ManagedManaged heap bytes. The secondary line shows the process working set.
GCGen 0, gen 1 and gen 2 collection counts.

The graphs

Three bar graphs show the retained history. They use fixed vertical scales so different runs are comparable:

GraphScaleUse it to spot
Frame time33 ms full heightOverall hitches and frames that missed the display budget.
Layout time16 ms full heightLayout rebuilds caused by state, styles or viewport changes.
Render time16 ms full heightDisplay-list submission and GPU-facing work.

A tall frame bar with short layout and render bars points at work outside the measured passes, such as waiting on the platform or another thread. A tall layout bar points at geometry churn. A tall render bar points at display-list or submission cost.

Frame budgets

Refresh rateBudget per frame
60 Hz16.7 ms
120 Hz8.3 ms
144 Hz6.9 ms

The graph scales are diagnostic aids, not hard budgets. A 16 ms layout bar on a 144 Hz display is already too large for a smooth frame.

A practical workflow

  1. Open the Performance tab and leave it visible.
  2. Reproduce one interaction, then stop. Avoid a long sequence that mixes several causes.
  3. Compare the three graphs. Is the spike in layout, render, or only frame time?
  4. Switch to Elements and inspect the subtree involved. Check for unnecessary class changes, large text updates, or layout-affecting styles.
  5. Reduce the work: keep subtrees stable, key list items, memoise derived data, and avoid triggering full-tree changes for local state.
Look for repeated spikes One tall bar is often startup or a one-off allocation. A regular pattern of spikes is the useful signal: it maps to an interaction or animation that runs on a timer or every frame.

What it does not tell you