[tor-relays] Consensus weight dropped

Roger Dingledine arma at mit.edu
Tue Jan 20 21:58:24 UTC 2015

On Tue, Jan 20, 2015 at 11:44:46AM +0100, Network Operations Center wrote:
> Thank you!
> https://atlas.torproject.org/#details/3D7E274A87D9A89AF064C13D1EE4CA1F184F2600

The votes from the directory authorities for the last consensus period
are here:

In this case it looks like schokomilch has these votes for the w line:

w Bandwidth=2525 Measured=1600  [moria1]
w Bandwidth=2525                [dizum]
w Bandwidth=2525                [Faravahar]
w Bandwidth=2525                [gabelmoo]
w Bandwidth=2525                [dannenberg]
w Bandwidth=2525                [urras]
w Bandwidth=2525                [longclaw]
w Bandwidth=2525 Measured=674   [tor26]
w Bandwidth=2525                [maatuska]

So since only two directory authorities vote a Measured value for it,
and the design calls for three opinions, it ends up unmeasured, and thus
with a consensus weight of 20.

You can read about the reasoning for requiring Measured votes here:

In theory gabelmoo and longclaw are supposed to have opinions about
your relay too:

But they don't, so here we are.

The problem is likely that the bwauth (bandwidth
authority) scripts are old and buggy and unmaintained. See
especially the tickets towards the bottom.

We've already known about this in the context of "the bandwidth
authority scripts are very poorly tuned for the changes that have
happened in the Tor network since the scripts were written, so they
vote wildly varying numbers for relays". But I don't think that
we'd realized the "some relays don't get three votes at all, so they
basically get zeroed out" issue. Hm.

(Ultimately I am hoping for the bwauth scripts to get phased out, in
favor of one of the secure bandwidth measurement schemes that various
research groups have been working on lately. Those other designs also
will have the advantage that it's harder to game the system by lying
about your bandwidth. But it will be some months at least until we have
one of those designs to evaluate.)


More information about the tor-relays mailing list