Performance/MemShrink/Meetings/2011-09-20

From MozillaWiki
Jump to: navigation, search
  • The 9/6 meeting was spent on triage, so there are no notes. There was no meeting on 9/13 due to the Mozilla All-Hands.
  • What is Tp5 (RSS)? Bhackett wants to know. Endurance tests are a better measure? 5% regression. Kyle or Justin Lebar will follow up with him. Alice may know about Talos stuff and what they are measuring.
  • Bug 686795: Kind of turning into a tracking bug for a certain kind of weird pages. Need more reporters, like text runs. Could we compare the number of each kind of DOM elements to what Chrome does and see if there are weird differences there?
  • Bug 686713: tons of memory for parent process. Kyle has noticed leaks running desktop on mobile, and devs don't run debug, and it isn't run on Tinderbox, so these could be real leaks. Alon will look into it.

MemShrink test day

  • Anthony (QA) is planning to do a MemShrink test day on Friday. What could they do? Going forward, they are planning on doing it once a month if that would be useful.
  • Web page for the test day.
  • Looking for zombie compartments.
  • Also could look for large differences in memory usage between Firefox and Chrome. What does "large" mean?
  • Ehsan was thinking people could look at how memory usage changes when people do things.
  • Another idea was comparing explicit memory usage before and after page load, but Justin said there's too much noise in the measurements for it to be useful.
  • How do we pick out new information from the noise of what we already know?
  • Browsing in parallel on two different browsers, carrying out similar actions? Set up a recording proxy? Get one person to do both, to avoid differences per account. Chrome has an about:memory of a sort.
  • How to do it in IE? Some kind of process explorer shows nice graphs. May want to limit to browsers where it is easy to deal with.
  • Goal is not necessarily to file bugs, but to gather data. Johnny and Justin will follow up with Anthony and will try to figure out.