[tor-relays] exit relay consensus weight

[Roger Dingledine]

On Thu, May 25, 2017 at 08:20:16PM -0700, Arisbe wrote:
> I just made an interesting observation that I thought I would share.
> Yesterday I started a VPS exit relay at a well known hosting company
> in Moldova [0]. Within 24 hours I saw the consensus weight exceed
> 10000.  The relay is bandwidth limited to 10 MiB/s.  Not that I'm
> complaining!

Thanks for running an exit relay!

(Using data files from
https://collector.torproject.org/recent/relay-descriptors/consensuses/)

$ grep -A4 "^r TorExitMoldova" 2017-05*|grep "w "
2017-05-24-20-00-00-consensus-w Bandwidth=0 Unmeasured=1
2017-05-24-21-00-00-consensus-w Bandwidth=20 Unmeasured=1
2017-05-24-22-00-00-consensus-w Bandwidth=20 Unmeasured=1
2017-05-24-23-00-00-consensus-w Bandwidth=20 Unmeasured=1
2017-05-25-00-00-00-consensus-w Bandwidth=20 Unmeasured=1
2017-05-25-01-00-00-consensus-w Bandwidth=20 Unmeasured=1
2017-05-25-02-00-00-consensus-w Bandwidth=20 Unmeasured=1
2017-05-25-03-00-00-consensus-w Bandwidth=20 Unmeasured=1
2017-05-25-04-00-00-consensus-w Bandwidth=20 Unmeasured=1
2017-05-25-05-00-00-consensus-w Bandwidth=20 Unmeasured=1
2017-05-25-06-00-00-consensus-w Bandwidth=20 Unmeasured=1
2017-05-25-07-00-00-consensus-w Bandwidth=20 Unmeasured=1
2017-05-25-08-00-00-consensus-w Bandwidth=20 Unmeasured=1
2017-05-25-09-00-00-consensus-w Bandwidth=20 Unmeasured=1
2017-05-25-10-00-00-consensus-w Bandwidth=8460
2017-05-25-11-00-00-consensus-w Bandwidth=8460
2017-05-25-12-00-00-consensus-w Bandwidth=8180
2017-05-25-13-00-00-consensus-w Bandwidth=8180
2017-05-25-14-00-00-consensus-w Bandwidth=8180
2017-05-25-15-00-00-consensus-w Bandwidth=8180
2017-05-25-16-00-00-consensus-w Bandwidth=8180
2017-05-25-17-00-00-consensus-w Bandwidth=8180
2017-05-25-18-00-00-consensus-w Bandwidth=8180
2017-05-25-19-00-00-consensus-w Bandwidth=9670
2017-05-25-20-00-00-consensus-w Bandwidth=9670
2017-05-25-21-00-00-consensus-w Bandwidth=9670
2017-05-25-22-00-00-consensus-w Bandwidth=9670
2017-05-25-23-00-00-consensus-w Bandwidth=9670
2017-05-26-00-00-00-consensus-w Bandwidth=9670
2017-05-26-01-00-00-consensus-w Bandwidth=9670
2017-05-26-02-00-00-consensus-w Bandwidth=9670
2017-05-26-03-00-00-consensus-w Bandwidth=9670

Here's what I think happened:

A) You started up your exit relay the evening of May 24 UTC, and it
published a descriptor with a tiny amount of bandwidth in it (from
self-testing).

B) Somehow, it attracted a traffic flow that was very high volume.
Its consensus weight was tiny, but there are millions of Tor clients,
so maybe one of them chose it anyway. Or maybe the bandwidth authorities
themselves added this load. I'm not sure how step 'B' happened, but
however it did, your relay handled a lot of traffic, so it learned that
it *could* handle a lot of traffic, so it published new relay descriptors
saying that it's quite fast.

It has published three descriptors so far. The third number on the
bandwidth line is its self-reported capacity:

published 2017-05-24 19:50:38
bandwidth 10485760 12582912 145408

published 2017-05-24 23:34:24
bandwidth 10485760 12582912 6487186

published 2017-05-25 15:39:43
bandwidth 10485760 12582912 11526593

C) By the time the bandwidth authorities got around to measuring it,
it was already proudly self-reporting a big capacity. The way the
bandwidth authorities work is that they decide a modification to the
self-reported number, depending on how you perform compared to your peers
who self-report a similar number. You perform about average compared
to your peers, so they gave you a consensus weight that is around the
number you were self-reporting.

> So it begs the question:  Is there not enough exit relays on the Tor
> network?

Well, exit relays attract traffic in a very different pattern than
guard relays. The blog post that we always point people to:
https://blog.torproject.org/blog/lifecycle-of-a-new-relay
has to do with how a fast non-exit relay will grow over time.

So it is much more normal for your consensus weight to grow quickly for
an exit relay. (Well, expected. It's hard to say what is normal with
the weird broken design that is the bandwidth authority subsystem these
days. :)

As for whether there are not enough exit relays... always! :) We actually
have about a third of the capacity of the network in Exits right now,
so from a load balancing perspective, it's not a disaster, since clients
avoid using the exit relays for any other positions in their circuits,
and it works out ok. But from the perspective of resistance against
correlation attacks, which is largely a function of diversity of entry
points and exit points, then having only 1/3 of the network as a possible
exit point means things aren't as good as they could be.

Hope that helps,
--Roger