Re: Summary of e10s performance (Talos + Telemetry + crash-stats)
For the e10s talos regressions see https://bugzilla.mozilla.org/show_bug.cgi?id=1174776 and https://bugzilla.mozilla.org/show_bug.cgi?id=1184277. We've already diagnose one source of the regression to be a difference with GC/CC behavior when running e10s talos. On Fri, Jul 10, 2015 at 5:44 PM, Vladan Djeric vdje...@mozilla.com wrote: Yup, the median shutdown duration for Release 39 users on Windows with Telemetry is 2.3 seconds for example: http://mzl.la/1HSHiD8 Those are also the kinds of shutdown times I see on my Windows machines when I have 3-5 windows open with 5-10 tabs each. What is your experience? Btw, you can go to about:telemetry and look through your archived Telemetry pings to see a history of your own shutdownDurations. Open about:telemetry, select Archived ping data, open the Simple Measurements section, and use the next-previous arrows to look through your Telemetry submissions. Focus on the saved-session pings. On Fri, Jul 10, 2015 at 5:33 PM, Mike Hommey m...@glandium.org wrote: On Fri, Jul 10, 2015 at 03:59:43PM -0400, Vladan Djeric wrote: A few of us on the perf team (+ Joel Maher) looked at e10s performance stability using Talos, Telemetry, and crash-stats. I wrote up the conclusions below. Notable improvements in Talos tests [1]: * Hot startup time in Talos improved by about 50% across all platforms (ts_paint [2]). This test measures time from Firefox launch until a Firefox window is first painted (ts_paint); I/O read costs are not accounted for, as data is already cached in the OS disk buffer before the test. * The tsvgr_opacity test improved 50-80% across all platforms. This is a sign of a reduction in the overhead of loading a page, instead of an improvement in actual SVG performance. * Linux scrolling performance improved 5-15% * The long-standing e10s WebGL performance regression has been fixed * SVG rendering performance (tsvgx) is ~25% better on Windows 7 8, but it is 10% worse on Windows XP and 25% worse on Linux Notable regressions in Talos tests [1]: * There are several large regressions unique to Windows XP. Scrolling smoothness regressed significantly (5-6 times worse on tp5o_scroll and tscrollx [2]), resizing of Firefox windows is 150% worse (tresize), SVG rendering performance is 25% worse (tsvgx) * Page loading time regressed across all platforms (tp5o). Linux regressed ~30%, OS X 10.10 regressed 20%, WinXP/Win8/Win7 all regressed ~10%. Page-loading with accessibility enabled (a11yr) saw similar regressions. * Time to open a new Firefox window (tpaint) regressed 30% on Linux, and across different versions of Windows (10%) * Resizing of Firefox windows (tresize) is ~15% worse on Linux * Note: not all tests are compatible with e10s yet (e.g. session-restore performance test) so this list isn't complete Notable improvements from Telemetry data [3]: * Overall tab animation smoothness improved significantly: 50% vs 30% of tab animation frames are hitting the target 16ms inter-frame interval. See FX_TAB_ANIM_* graphs in [3] to see the distribution of frame intervals. Note that not all tab animations benefited equally. * e10s significantly decreased jank caused by GC CC, both in parent content processes (GC_MAX_PAUSE_MS, GC_SCC_SWEEP_MAX_PAUSE_MS, CYCLE_COLLECTOR_MAX_PAUSE, etc [3]) * Unlike Talos, Telemetry suggests that the time to open a new Firefox window improved with e10s (FX_NEW_WINDOW_MS) * Median time to restore a saved session improved by 40ms or 20% (simpleMeasurements/sessionRestored) * Median shutdown duration improved by 120ms or 10% (simpleMeasurements/shutdownDuration) Wait. What? Median shutdown duration is 1.2s ?!? ___ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform ___ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
Summary of e10s performance (Talos + Telemetry + crash-stats)
A few of us on the perf team (+ Joel Maher) looked at e10s performance stability using Talos, Telemetry, and crash-stats. I wrote up the conclusions below. Notable improvements in Talos tests [1]: * Hot startup time in Talos improved by about 50% across all platforms (ts_paint [2]). This test measures time from Firefox launch until a Firefox window is first painted (ts_paint); I/O read costs are not accounted for, as data is already cached in the OS disk buffer before the test. * The tsvgr_opacity test improved 50-80% across all platforms. This is a sign of a reduction in the overhead of loading a page, instead of an improvement in actual SVG performance. * Linux scrolling performance improved 5-15% * The long-standing e10s WebGL performance regression has been fixed * SVG rendering performance (tsvgx) is ~25% better on Windows 7 8, but it is 10% worse on Windows XP and 25% worse on Linux Notable regressions in Talos tests [1]: * There are several large regressions unique to Windows XP. Scrolling smoothness regressed significantly (5-6 times worse on tp5o_scroll and tscrollx [2]), resizing of Firefox windows is 150% worse (tresize), SVG rendering performance is 25% worse (tsvgx) * Page loading time regressed across all platforms (tp5o). Linux regressed ~30%, OS X 10.10 regressed 20%, WinXP/Win8/Win7 all regressed ~10%. Page-loading with accessibility enabled (a11yr) saw similar regressions. * Time to open a new Firefox window (tpaint) regressed 30% on Linux, and across different versions of Windows (10%) * Resizing of Firefox windows (tresize) is ~15% worse on Linux * Note: not all tests are compatible with e10s yet (e.g. session-restore performance test) so this list isn't complete Notable improvements from Telemetry data [3]: * Overall tab animation smoothness improved significantly: 50% vs 30% of tab animation frames are hitting the target 16ms inter-frame interval. See FX_TAB_ANIM_* graphs in [3] to see the distribution of frame intervals. Note that not all tab animations benefited equally. * e10s significantly decreased jank caused by GC CC, both in parent content processes (GC_MAX_PAUSE_MS, GC_SCC_SWEEP_MAX_PAUSE_MS, CYCLE_COLLECTOR_MAX_PAUSE, etc [3]) * Unlike Talos, Telemetry suggests that the time to open a new Firefox window improved with e10s (FX_NEW_WINDOW_MS) * Median time to restore a saved session improved by 40ms or 20% (simpleMeasurements/sessionRestored) * Median shutdown duration improved by 120ms or 10% (simpleMeasurements/shutdownDuration) Notable regressions from Telemetry data [3]: * Unlike Talos, Telemetry numbers imply that the median real-world startup time, measured as time to first-paint, regressed by 550ms or 20% with e10s (simpleMeasurements/firstPaint) * The frequency of jank events lasting more than 100ms increased from ~19 events/min to ~21 events/minute with e10s. This was derived from the main-thread's event processing times and session uptime (gecko_hangs_per_minute) * Similarly the frequency of the slow-script dialog appearing seems to have roughly doubled with e10s (histograms/SLOW_SCRIPT_NOTICE_COUNT) * A side-note: interpreting Telemetry data is trickier than Talos data, because Telemetry measurements aren't gathered from a controlled environment, there are confounding variables, opt-in bias, etc. An additional challenge with e10s Telemetry is that many measurements haven't yet been re-validated to confirm that they measure the same things in e10s and non-e10s. Notable stability improvements [4]: * E10S Firefox can survive crashes in content-process code, so it's no surprise that the E10S parent process crash rate is a quarter of the single-process crash rate (based on crash-stats from Nightly 42 on Windows [4]) * The total number of E10S crashes of any type (content crash or parent crash) is roughly the same as without E10S * Oddly enough, E10S seems to win on plugin crash rates as well! [4] * There seem to be no regressions in crash rate compared to single-process References: 1. Joel Maher used compare-talos to compare the Talos scores of an m-c revision (Nightly 42) in e10s non-e10s configurations: https://bugzilla.mozilla.org/show_bug.cgi?id=1144120#c5 Data in friendlier chart form: https://drive.google.com/open?id=1qfkcoE5_25GtZDa-pIlqFMw6pLueplOsP1YhQb8UcC8 Talos data was gathered from a Firefox 42 m-c build aad95360a002 from June 29th 2. Talos test descriptions https://wiki.mozilla.org/Buildbot/Talos/Tests 3. Roberto Vitillo compared Telemetry from 150,000 Nightly sessions submitted on June 15th with buildIDs in the range [20150601, 20150616]: http://nbviewer.ipython.org/urls/gist.githubusercontent.com/vitillo/cb6f1304316c1c1a2cbc/raw/e10s%20analysis.ipynb ~90% of the sessions were from e10s clients, so interpret the e10s non-e10s populations with a grain of salt. The numbers were not broken down by OS. I did not comment on any findings where the delta had more than 0.10 probability of being caused by chance. 4. Crash-stats comparisons from
Re: Summary of e10s performance (Talos + Telemetry + crash-stats)
On Fri, Jul 10, 2015 at 03:59:43PM -0400, Vladan Djeric wrote: A few of us on the perf team (+ Joel Maher) looked at e10s performance stability using Talos, Telemetry, and crash-stats. I wrote up the conclusions below. Notable improvements in Talos tests [1]: * Hot startup time in Talos improved by about 50% across all platforms (ts_paint [2]). This test measures time from Firefox launch until a Firefox window is first painted (ts_paint); I/O read costs are not accounted for, as data is already cached in the OS disk buffer before the test. * The tsvgr_opacity test improved 50-80% across all platforms. This is a sign of a reduction in the overhead of loading a page, instead of an improvement in actual SVG performance. * Linux scrolling performance improved 5-15% * The long-standing e10s WebGL performance regression has been fixed * SVG rendering performance (tsvgx) is ~25% better on Windows 7 8, but it is 10% worse on Windows XP and 25% worse on Linux Notable regressions in Talos tests [1]: * There are several large regressions unique to Windows XP. Scrolling smoothness regressed significantly (5-6 times worse on tp5o_scroll and tscrollx [2]), resizing of Firefox windows is 150% worse (tresize), SVG rendering performance is 25% worse (tsvgx) * Page loading time regressed across all platforms (tp5o). Linux regressed ~30%, OS X 10.10 regressed 20%, WinXP/Win8/Win7 all regressed ~10%. Page-loading with accessibility enabled (a11yr) saw similar regressions. * Time to open a new Firefox window (tpaint) regressed 30% on Linux, and across different versions of Windows (10%) * Resizing of Firefox windows (tresize) is ~15% worse on Linux * Note: not all tests are compatible with e10s yet (e.g. session-restore performance test) so this list isn't complete Notable improvements from Telemetry data [3]: * Overall tab animation smoothness improved significantly: 50% vs 30% of tab animation frames are hitting the target 16ms inter-frame interval. See FX_TAB_ANIM_* graphs in [3] to see the distribution of frame intervals. Note that not all tab animations benefited equally. * e10s significantly decreased jank caused by GC CC, both in parent content processes (GC_MAX_PAUSE_MS, GC_SCC_SWEEP_MAX_PAUSE_MS, CYCLE_COLLECTOR_MAX_PAUSE, etc [3]) * Unlike Talos, Telemetry suggests that the time to open a new Firefox window improved with e10s (FX_NEW_WINDOW_MS) * Median time to restore a saved session improved by 40ms or 20% (simpleMeasurements/sessionRestored) * Median shutdown duration improved by 120ms or 10% (simpleMeasurements/shutdownDuration) Wait. What? Median shutdown duration is 1.2s ?!? ___ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
Re: Summary of e10s performance (Talos + Telemetry + crash-stats)
Yup, the median shutdown duration for Release 39 users on Windows with Telemetry is 2.3 seconds for example: http://mzl.la/1HSHiD8 Those are also the kinds of shutdown times I see on my Windows machines when I have 3-5 windows open with 5-10 tabs each. What is your experience? Btw, you can go to about:telemetry and look through your archived Telemetry pings to see a history of your own shutdownDurations. Open about:telemetry, select Archived ping data, open the Simple Measurements section, and use the next-previous arrows to look through your Telemetry submissions. Focus on the saved-session pings. On Fri, Jul 10, 2015 at 5:33 PM, Mike Hommey m...@glandium.org wrote: On Fri, Jul 10, 2015 at 03:59:43PM -0400, Vladan Djeric wrote: A few of us on the perf team (+ Joel Maher) looked at e10s performance stability using Talos, Telemetry, and crash-stats. I wrote up the conclusions below. Notable improvements in Talos tests [1]: * Hot startup time in Talos improved by about 50% across all platforms (ts_paint [2]). This test measures time from Firefox launch until a Firefox window is first painted (ts_paint); I/O read costs are not accounted for, as data is already cached in the OS disk buffer before the test. * The tsvgr_opacity test improved 50-80% across all platforms. This is a sign of a reduction in the overhead of loading a page, instead of an improvement in actual SVG performance. * Linux scrolling performance improved 5-15% * The long-standing e10s WebGL performance regression has been fixed * SVG rendering performance (tsvgx) is ~25% better on Windows 7 8, but it is 10% worse on Windows XP and 25% worse on Linux Notable regressions in Talos tests [1]: * There are several large regressions unique to Windows XP. Scrolling smoothness regressed significantly (5-6 times worse on tp5o_scroll and tscrollx [2]), resizing of Firefox windows is 150% worse (tresize), SVG rendering performance is 25% worse (tsvgx) * Page loading time regressed across all platforms (tp5o). Linux regressed ~30%, OS X 10.10 regressed 20%, WinXP/Win8/Win7 all regressed ~10%. Page-loading with accessibility enabled (a11yr) saw similar regressions. * Time to open a new Firefox window (tpaint) regressed 30% on Linux, and across different versions of Windows (10%) * Resizing of Firefox windows (tresize) is ~15% worse on Linux * Note: not all tests are compatible with e10s yet (e.g. session-restore performance test) so this list isn't complete Notable improvements from Telemetry data [3]: * Overall tab animation smoothness improved significantly: 50% vs 30% of tab animation frames are hitting the target 16ms inter-frame interval. See FX_TAB_ANIM_* graphs in [3] to see the distribution of frame intervals. Note that not all tab animations benefited equally. * e10s significantly decreased jank caused by GC CC, both in parent content processes (GC_MAX_PAUSE_MS, GC_SCC_SWEEP_MAX_PAUSE_MS, CYCLE_COLLECTOR_MAX_PAUSE, etc [3]) * Unlike Talos, Telemetry suggests that the time to open a new Firefox window improved with e10s (FX_NEW_WINDOW_MS) * Median time to restore a saved session improved by 40ms or 20% (simpleMeasurements/sessionRestored) * Median shutdown duration improved by 120ms or 10% (simpleMeasurements/shutdownDuration) Wait. What? Median shutdown duration is 1.2s ?!? ___ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform