[Author Prev][Author Next][Thread Prev][Thread Next][Author Index][Thread Index]

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



Hi all,

Wanted to provide an update (even if it's not as good news as I hoped
to give) because I know this is a very frustrating issue for everyone.

At a high level, the bwauth scripts segment the network into four
segments ranked by relay speed, and measure each of these segments.
They are 0-10, 10-40[0], 30-60, 60-100 (so the top 10% of relays by
speed, and so on.)  After 26 hours[1], the first scanner we added is
currently at:

Segment 1: Completed, looped a couple times
Segment 2: 34.2 of the 10-40
Segment 3: 46.5 of the 30-60
Segment 4: 77.4 out of the 60-100

This isn't very encouraging - it's slow going.  Two of things that
were just brainstormed in #tor-dev were:

1) Increasing the number of scanners tended to overwhelm the tor
instance that supported these scanners, so we want to try doubling the
scanners *and* the tors. Hopefully this will let us work our way
through the list of relays faster.
2) We want to look at the possibility of relays moving around in the
percentiles in each consensus, getting unlucky, and not being
measured; potentially fixable by pinning a relay to a percentile, and
then when we get all the way through a segment, unpin it, get that
segment from the latest consensus, and restart.  This may result in a
relay being pinned into scanner#2, scanner#1 completes, measures the
relay, then scanner#2 measures it... that wouldn't be so bad, double
measuring is better than not measuring. When torflow was written over
four years ago, there may have been a good reason it didn't work this
way, and we need to see if we can re-reason out what that was.

So we are working on identifying the subtler issues and seeing if we
can fix them, in addition to just adding more of the same.

-tom [2]


[0] The 10-40 overlap is a bug we just found.
[1] Well, more like 50 hours, but the first 24 were lost because of a
breaking change in a python point release
[2] (Most of this is aagbsn's knowledge, I'm just transcribing it.)
_______________________________________________
tor-relays mailing list
tor-relays@xxxxxxxxxxxxxxxxxxxx
https://lists.torproject.org/cgi-bin/mailman/listinfo/tor-relays