connection timeouts are affecting Tor Browser usability

I have spent some time watching circuit and stream events while connecting to different sites. I telnet into tor's config port using the following command (using ts to give time stamps):

telnet localhost 9151 | ts

I open the browser console and get the tor password by entering m_tb_control_pass And then I paste the result like this: authenticate [value of m_tb_control_pass] Finally I enter setevents circ stream.

What I have noticed is that a significant fraction of new site connections result in at least 1 timeout of 10 seconds. (Tor Browser's CircuitStreamTimeout is set to 0, which results in a timeout equal to MIN_CIRCUIT_STREAM_TIMEOUT, or 10 seconds.) Here's what a timeout looks like:

Feb 03 19:00:03 650 STREAM 868 NEW 0 people.torproject.org:443 SOURCE_ADDR=127.0.0.1:50318 PURPOSE=USER
Feb 03 19:00:03 650 CIRC 149 LAUNCHED BUILD_FLAGS=NEED_CAPACITY PURPOSE=GENERAL TIME_CREATED=2017-02-04T03:00:03.738597
Feb 03 19:00:03 650 CIRC 149 EXTENDED [...] BUILD_FLAGS=NEED_CAPACITY PURPOSE=GENERAL TIME_CREATED=2017-02-04T03:00:03.738597 SOCKS_USERNAME="torproject.org" SOCKS_PASSWORD="7d8ea4ccf4ba6345846e0fccacd4d941"
Feb 03 19:00:04 650 CIRC 149 EXTENDED [...] BUILD_FLAGS=NEED_CAPACITY PURPOSE=GENERAL TIME_CREATED=2017-02-04T03:00:03.738597 SOCKS_USERNAME="torproject.org" SOCKS_PASSWORD="7d8ea4ccf4ba6345846e0fccacd4d941"
Feb 03 19:00:04 650 CIRC 149 EXTENDED [...] BUILD_FLAGS=NEED_CAPACITY PURPOSE=GENERAL TIME_CREATED=2017-02-04T03:00:03.738597 SOCKS_USERNAME="torproject.org" SOCKS_PASSWORD="7d8ea4ccf4ba6345846e0fccacd4d941"
Feb 03 19:00:04 650 CIRC 149 BUILT [...] BUILD_FLAGS=NEED_CAPACITY PURPOSE=GENERAL TIME_CREATED=2017-02-04T03:00:03.738597 SOCKS_USERNAME="torproject.org" SOCKS_PASSWORD="7d8ea4ccf4ba6345846e0fccacd4d941"
Feb 03 19:00:04 650 STREAM 868 SENTCONNECT 149 people.torproject.org:443
[...]
Feb 03 19:00:14 650 STREAM 868 DETACHED 149 people.torproject.org:443 REASON=TIMEOUT
Feb 03 19:00:14 650 CIRC 150 LAUNCHED BUILD_FLAGS=NEED_CAPACITY PURPOSE=GENERAL TIME_CREATED=2017-02-04T03:00:14.588714

After a timeout occurs, the tor client closes the circuit, builds a new circuit and attempts to connect to the same site again. This repeats at least 3 times.

I did an experiment where I connected to https://people.torproject.org/~arthuredelstein (a page with hardly any content) and then repeatedly selected "New Tor Circuit for this Site" 50 times.

Here are the results for 50 reloads. Each digit represents the number of 10-second stream timeouts observed before a given connection succeeded. 20020000000000000000002010000000000001000100000103

In other words 8 out of 50 connections showed a timeout. And interestingly, four of these connections exhibited a double or triple timeout (20 or 30 seconds delay).

I think this may be a big part of the perception of Tor Browser as "slow". Actual loading of pages doesn't seem drastically slow to me, and once I have successfully connected to a new site, following links to other pages on the same site (i.e., the same circuit) is usually acceptable.

(I also did another quick test on another site and 5/25 connections had at least 1 timeout.)

So here are some questions for further investigation:

• Why are there so many timeouts? Are any of these timeouts due to silent errors in a Tor node? (If such errors could be promptly reported back to the client, maybe we could avoid the waiting for the long timeout.)
• What's the reason for MIN_CIRCUIT_STREAM_TIMEOUT being 10 seconds? Would it do any harm to make this shorter, say 5 seconds or 2 seconds?
• So many double or triple timeouts are suspicious, because each timeout in a double or triple is reported for a different circuit. Could this mean the connection error is caused by the client or guard rather than a connection failure at the exit node?