[tor-relays] tor network "loses" ~50 relays/day due to bw auth problem

[Roger Dingledine]

On Mon, Jun 01, 2015 at 11:02:53AM -0500, Tom Ritter wrote:
> Hrm.  So this gets into the inner workings of the bwauth system which
> is... complicated.[0] Honestly, I'm not actually sure how the
> individual data from the different bwauths is combined into a single
> value for the consensus.

The weight in the consensus is the median of the bwauth votes.

(That design aims to best resist manipulation of the weight by a small
number of bwauths.)

> I'm not sure what the answer is for your problem, but I'm beginning to
> wonder if the general approach to this problem is "There should be a
> bwauth debugging mechanism similar to
> https://consensus-health.torproject.org/ or (the unimplemented)
> proposal 164."   I don't know if said mechanism would be a component
> of atlas/globe or an entirely separate site, but if the bwauths
> exported their hourly files, and this hypothetical tool aggregated
> that data, it may help figure these things out.  It might give an
> answer like "moria spazzed out and undervoted me, let me wait until it
> scans me again."

Yes, I'd love to have something like this. Once upon a time I asked
for us to export more of the bwauth internals:
https://trac.torproject.org/projects/tor/ticket/2532
But at the time Mike decided that the intermediate bwauth numbers weren't
useful enough to export.

Another ticket you might enjoy is:
https://trac.torproject.org/projects/tor/ticket/2394

I think ultimately we need somebody to simplify the bwauth design as
much as possible (but no more), and then we can start assessing its
output to see how it compares to reality.

In parallel, or maybe instead, we might want more work on bandwidth
estimation algorithms that are resilient to cheating. Right now the bwauth
results can be gamed, and making them harder to game is an open research
area. Some research groups have been working on it, and I look forward to
some of their papers going public. But none of the papers that I've seen
so far are perfect designs, so don't set your expectations too high. :)

But all of that said, just visualizing the various Measured= lines by
each of the bwauths might be a great start, and this can be done with
the published votes as-is:
http://freehaven.net/~arma/moria1-v3-status-votes

--Roger