Break Wed and Wee weights into two classes each

[mikeperry]

In 0.2.2, we added consensus bandwidth-weights to specify how to distribute node selection choices for the three positions in a circuit among the different node flags (Guard, Exit, Guard+Exit and None). See 3.4.3 of
​https://gitweb.torproject.org/tor.git/blob_plain/HEAD:/doc/spec/dir-spec.txt
and
​http://archives.seul.org/or/dev/Sep-2010/msg00016.html. See also the original post:
​http://archives.seul.org/or/dev/Jan-2010/msg00012.html

I added some code to statsplitter.py to display the breakdown of bytes by port, using Karsten's extra-info exit ports stats:

Default exit blutmagie read 503.9M

other: 56.6% 80: 38.9% 51413: 1.4% 443: 0.7% 4000: 0.6% 49354: 0.5%

Default exit blutmagie wrote 182.6M

other: 92.1% 80: 4.0% 51413: 1.6% 24273: 1.2% 53163: 0.4% 443: 0.3%

Misc Exit Amunet3 read 206.1M

80: 97.1% 443: 1.6% 8080: 0.6% 81: 0.3% 563: 0.2% other: 0.1%

Misc Exit Amunet3 wrote 7.5M

80: 88.0% 443: 9.0% 8080: 1.3% other: 1.3% 8888: 0.2% 81: 0.1%

Amunet has the non-default exit policy listed in #6 at:
​https://blog.torproject.org/blog/tips-running-exit-node-minimal-harassment

Blutmagie has the default.

So what this tells us is that we may want to devote circuits predicted to be used for port 80 or 443 more on the non-default exit policy than for the default exit policy to compensate for the fact that these non-default exits with normally otherwise get less traffic than their default exit policy peers.

[mikeperry]

One wrinkle that makes this less optimal/more difficult is that we usually are only building circuits with a *guess* as to what the port is, at best, during prediction. It is rare that we're building a circuit for a specific stream...

[arma]

Assigning to 0.2.3.x so it has a milestone.

Mike, how come this ticket doesn't have a point estimate? :)

[mikeperry]

Because we're still not sure if we even want to do it. Maybe this becomes less relevant with guard bw throttling?