[metrics-bugs] #34303 [Metrics/Onionperf]: Find out why onion service measurements have gotten slower

[Tor Bug Tracker & Wiki]

#34303: Find out why onion service measurements have gotten slower
-------------------------------+--------------------------------
 Reporter:  karsten            |          Owner:  metrics-team
     Type:  defect             |         Status:  new
 Priority:  Medium             |      Milestone:
Component:  Metrics/Onionperf  |        Version:
 Severity:  Normal             |     Resolution:
 Keywords:                     |  Actual Points:
Parent ID:                     |         Points:
 Reviewer:                     |        Sponsor:  Sponsor59-must
-------------------------------+--------------------------------

Comment (by karsten):

 Replying to [comment:6 arma]:
 > In the graph above, it looks like the "before" and "after" overlap for a
 bit of time, at the end of April? So how come the old measurements are
 fast, and the new measurements are slow, even when they're done at the
 same time?

 Exactly, that's the question. The Tor '''network''' was the same, yet the
 old and new instances produced vastly different results.

 So, first results from `"Schedulers Vanilla"` are that this setting does
 not affect measurements at all:

 [[Image(onionperf-2020-05-27.png, 700px)]]

 Some observations:

  - The blue line is from the old `op-hk` instance from early April. It
 didn't run anymore towards the end of April, but judging from the graph
 above we would expect its performance to stay the same. Note that it shows
 time to first byte, not time to 50 KiB, though the difference is probably
 minor.

  - The orange line is from the new `op-hk2` instance from yesterday.

  - The green line is from the newly setup Hong Kong instance with the
 `"Schedulers Vanilla"` setting. I'm 90% certain that I did not mess up
 setting that option. There's no easy way to check, because we're not
 writing out the `torrc` file to disk but feeding it into the `tor`
 processes via stdin. However, when I first tried it, I passed `"Schedulers
 Vanilla"` without newline, and it complained about `"VanillaRunAsDaemon"`
 not being a valid scheduler option, and it stopped complaining after I
 passed `"Schedulers Vanilla\n"`. So, it might have worked. The results are
 equally bad as in the orange line.

  - Note that I made this graph using yet unmerged OnionPerf branches. The
 TTFB definition in this graph is the same as in the other graphs above.

 I'll try older Tor versions today. The `op-hk` instance was running,
 *ahem*, 0.2.9, whereas the new instances are running 0.3.5. I'll see what
 versions between those two I get compiled and running, and I might also
 try newer versions.

--
Ticket URL: <https://trac.torproject.org/projects/tor/ticket/34303#comment:7>
Tor Bug Tracker & Wiki <https://trac.torproject.org/>
The Tor Project: anonymity online