[tor-bugs] #22798 [Core Tor/Tor]: Windows relay is several times slower than Linux relay

[Tor Bug Tracker & Wiki]

#22798: Windows relay is several times slower than Linux relay
-------------------------------------------------+-------------------------
 Reporter:  Vort                                 |          Owner:  (none)
     Type:  defect                               |         Status:
                                                 |  needs_revision
 Priority:  Medium                               |      Milestone:  Tor:
                                                 |  0.3.3.x-final
Component:  Core Tor/Tor                         |        Version:  Tor:
                                                 |  0.2.9.11
 Severity:  Normal                               |     Resolution:
 Keywords:  tor-relay, performance, windows,     |  Actual Points:
  review-group-29                                |
Parent ID:                                       |         Points:  5
 Reviewer:                                       |        Sponsor:
-------------------------------------------------+-------------------------

Comment (by Vort):

 > What do you mean above when you say "It breaks build"?

 It was said only about v1 patch. Newer versions does not have such
 problem.

 > The function update_send_buffer_size should probably be static

 Yes, thanks. Changed it.

 > and it definitely needs documentation.
 > It should probably explain how it is doing that, and what exactly it is
 doing.

 My English is not good, so new description may need rewording.

 > I see above that you have done work on measuring the performance impact,
 both for the syscalls and for the network performance. Can you summarize
 the outcomes?

 According to [attachment:tor_sampling.png profiling results], `WSAIoctl()`
 and `setsockopt()` functions use 1% of CPU resources. But I am not sure
 about correctness of this result, since no API function is approaching 85%
 of CPU usage as mentioned in ticket #23433.
 Concerning network: in isolated environment, test program was able to
 accelerate to 4 MiB/s, which allows to think that there are no limit on
 upload speed after fix (or limit is high enough). Going back to real
 tests: without patch, upload speed for my relay on high latency link was
 50 KiB/s and BwAuth's weight was 20. With patch, all allocated bandwidth
 (3 MiB/s) is used by relay and weight is increased to 4000.

 >> Maybe non-encrypted connection requires update_send_buffer_size call
 too?
 > Yes, good idea.

 Are there any method to test upload speed with non-encrypted connections?
 And where that call should be placed? In `else` branch before `n_written =
 (size_t) result;` line?

--
Ticket URL: <https://trac.torproject.org/projects/tor/ticket/22798#comment:132>
Tor Bug Tracker & Wiki <https://trac.torproject.org/>
The Tor Project: anonymity online