Enable Show Tab Performance Monitor in the Debug menu, which Show
Debug menu in Settings → Advanced turns on, or set
debug.ui.show_tab_performance_monitor to true. It defaults to false.
A row of its own below the page displays the current tab's statistics.
Each tab retains its own CPU history while monitoring is enabled.
LibWebView::TabPerformanceMonitor resolves browser ownership and supplies
TabPerformanceStats to the frontend. TabPerformanceAccumulator performs
counter differencing and windowing independently of the UI and OS APIs.
- CPU: CPU-time deltas divided by actual monotonic elapsed time. 100% is one logical CPU; values above 100% are expected. All live and terminated threads in exclusively owned WebContent processes contribute, so a worker counts under the WebContent process that hosts it. Registered page and embedded-frame ownership identify WebContent owners. Processes shared across tabs remain unattributed, and so does a process that runs on without pages for its shared workers. The first sample establishes a baseline. The sparkline retains ten seconds.
- MEM: An estimate of owned process memory, counted once per process.
macOS uses the kernel's physical-footprint ledger, including compressed
memory, rather than RSS. Linux uses resident minus shared pages from
/proc/PID/statm, a cheap private-resident approximation that omits file-backed private pages and swapped-out memory. These platform estimates are not directly comparable. Other platforms report unavailable until a collector is implemented. Shared helper/cache allocations and resources without a process-level ownership charge are excluded. - Download/upload: HTTP response/request body bytes transferred by curl, before response decompression. Disk-cache hits contribute no traffic. Headers, TLS/TCP overhead, WebSockets, and helper-initiated background requests without an originating page or exclusive worker are excluded. A request keeps its originating process/page context when transferred to another client. The browser resolves that context to a tab; traffic whose original page or worker has already disappeared is left unattributed. Byte units are decimal. Rates use a rolling second, with delayed batches weighted by their actual sampling interval.
- FPS: Accepted compositor backing-store swaps over the last second,
including compositor-driven animation. This measures frames delivered to
the frontend, not hardware display scanout or
requestAnimationFramecalls. After a second without a frame the value isFPS —. FPS uses neutral text.
CPU, memory and FPS values use neutral text. Labels use 65% opacity and the sparkline uses a subdued stroke. Network download is blue and upload orange. Fixed-width columns and tabular digits prevent layout shifts.
The browser samples owned process counters every 500 ms. RequestServer runs its own 500 ms timer and pushes byte deltas asynchronously to the browser; there is no UI polling or synchronous traffic-statistics IPC. Completed requests contribute their final delta before their curl handle is released. Batches coalesce requests with the same originating process/page.
Disabling the setting destroys both timers, clears histories and pending byte batches, and disconnects the browser's traffic callback. RequestServer does not sample curl counters or push traffic messages while disabled. Existing request setup retains small ownership metadata, and the frame delivery path has a disabled-state check. Re-enabling establishes fresh CPU and in-flight request baselines; restarting RequestServer restores an enabled subscription.
Shared helper CPU is currently excluded: the network, image decoder, compiler and compositor processes do not yet expose per-job CPU accounting. Their whole-process CPU is never assigned to a tab. Additional explicitly owned memory or helper CPU can be added to the accounting layer without putting platform or ownership queries in toolbar widgets.