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

Re: Exit Balancing Patch

Thus spake Cat Okita (cat@xxxxxxxxxxxx):

> On Thu, 26 Jul 2007, Mike Perry wrote:
> >That is what load balancing is for though. If you notice, that is why
> >I did all my userbase size calculations based on a 10KB/sec stream
> >rate, which is the average speed of the rest of the Tor network. As I
> >pointed out, the top 5% has room for SEVEN TIMES MORE 10KB/sec
> >streams, since its average stream capacity is 70K/sec.
> I'm still working through this thread, but I'd like to point out that
> you're presuming that changing the amount of bandwidth available will
> automatically result in additional users, which will offset the fact
> that SEVEN TIMES MORE 10KB/sec streams will be going through the top 5%
> tor nodes - which to me, at least, suggests SEVEN TIMES MORE opportunities
> for compromise via the top 5% of nodes...

If this is the case, why don't we just route everything uniformly? It
was my impression we accepted that nodes with higher bandwidth have
greater potential for compromise, but that it was a smart thing to
route streams through them proportional to their capacity...

> Do you expect that adding SEVEN TIMES MORE 10KB/sec streams will actually
> result in SEVEN TIMES MORE users to offset this (and if that's the case,
> we're right back where we started, in terms of performance issues...)

Do you use Tor regularly? Do you talk to real people who try to do so?
Maybe you don't notice the problem because you got lucky enough to
have fast guard nodes. But those circuit failures, timeouts, and
usability issues in the lower guard nodes are a real problem. People
who get those guards find Tor nearly unusable - around 20% of their
streams timeout (after a solid minute of doing nothing), and another
20% randomly have to restart on a different circuit. As sjmurdoch
discussed on or-talk, there is a definite correlation to percieved
performance and adoption.

Once again, it is an equilibrium thing. Wherever the equilibrium
settles wrt user pain threshholds and bandwidth, that is where it

If the network is properly balanced, that high average capacity in the
top nodes should disappear, and we should be able to support more
users. Or maybe we won't have more users, maybe users will still
decide Tor is too slow for them and will walk away. But then we should
be left with more bandwidth. The point is, that bandwidth should be
evenly distributed across all streams and all nodes, in a properly
load balanced network.

Does this make sense?

Mike Perry
Mad Computer Scientist
fscked.org evil labs

Attachment: pgpxajZa6rMyR.pgp
Description: PGP signature