>> >> As Joel mentioned, it's pretty easy to schedule profiling runs for talos >> using trychooser. Scheduling a profiling run as part of the >> regression-filing process is something we could consider doing, if >> there's a broad consensus it would be useful (I'm always wary of putting >> extra load on the machines for something that would be only useful in >> 10%, say, of cases).
I think it would be very useful. Performance profiles are, imo, the best first step in attempting to drill down into the main issue causing the regression. > (I'm always wary of putting >>> extra load on the machines for something that would be only useful in >>> 10%, say, of cases). Note that for most cases you'd only need or want a single run (multiple runs with profiling aren't generally useful unless you're looking for an intermittent performance problem). You also wouldn't need to do a full rebuild if you could trigger the single profiling run on the same builds that we've been doing the retriggers on. On 19/01/2016 10:06 AM, William Lachance wrote: > On 2016-01-18 4:42 AM, Nicolas B. Pierron wrote: >> >> I agree, this should be the part of the developer to work that out, but >> the TS Paint benchmark is out of the knowledge base of JS developers. I >> feel that the problem is reaching developers with a wording known by the >> developers. > > So we have a wiki page which describes each Talos test in detail, we > link to these in all perf bugs, but it can be useful reading on its own. > For example, ts_paint is described here: > > https://wiki.mozilla.org/Buildbot/Talos/Tests#ts_paint > > I'm sure some of the test descriptions could use some improvements, feel > free to ask on #perf if anything's unclear. > >> This is just a raw idea, but maybe this would make more sense to provide >> a diff of profiles, and show what decreased / increased. At least this >> would make these benchmarks less obscure. > > As Joel mentioned, it's pretty easy to schedule profiling runs for talos > using trychooser. Scheduling a profiling run as part of the > regression-filing process is something we could consider doing, if > there's a broad consensus it would be useful (I'm always wary of putting > extra load on the machines for something that would be only useful in > 10%, say, of cases). > > Will > _______________________________________________ > 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
