Ran again on ToT (chromium: b7b1fc005998a4209e8979081305608e8e85c8ed).

time testing/xvfb.py tools/perf/run_benchmark v8.browsing_desktop --browser=exact --browser-executable=out/Release/chrome  --results-label=TOT  --story-filter=browse:news:nytimes$ --pageset-repeat=1  --show-stdout --reset-results 

with Perfetto:
real	5m54.444s
user	9m48.484s
sys	0m52.373s

without Perfetto:
real	4m25.603s
user	7m1.338s
sys	0m40.395s

That's a 25% regression.

I believe perfetto splits all trace events into two (start / end). The original backend could reduce TRACE_EVENT_PHASE_COMPLETE events into a single entry in the json, so the resulting json file with perfetto may be larger and thus slower to parse and analyze. Dunno if we can work around that somehow?

Is that the only possible reason for the benchmark script to have gotten slower?

📍 Pinpoint job started.
https://pinpoint-dot-chromeperf.appspot.com/job/1464ba20140000

Started a bisect on a huge regression visible here: https://chromeperf.appspot.com/report?sid=e004cacde6bdf79bede350d6ddaa6a806449b1c37e948afe086f719127a2ad1c&start_rev=606619&end_rev=609968

I thought that was Perfetto but cannot see it in the logs; maybe I was just mistaken... 

I actually sped up the metrics computation by rolling V8 from M57 to M70 but it still feels very slow.

📍 Couldn't reproduce a difference.
https://pinpoint-dot-chromeperf.appspot.com/job/1464ba20140000

📍 Pinpoint job started.
https://pinpoint-dot-chromeperf.appspot.com/job/12105b0fe40000

📍 Couldn't reproduce a difference.
https://pinpoint-dot-chromeperf.appspot.com/job/12105b0fe40000

We discussed a few alternatives to splitting the _COMPLETE events, but none seem super feasible. Switching from JSON to proto in the future should more than compensate for this, though.

 mlippautz: Can you retry this after http://crrev.com/612360 please?

Happy to check once it has relanded. Thanks for following up on these issues!