Exit Balancing Patch

[Mike Perry]

Thus spake Roger Dingledine (arma at mit.edu):

> On Wed, Jul 18, 2007 at 02:18:56AM -0700, Mike Perry wrote:
> > Why does it matter if routers are CPU or network bound? From the Tor
> > network routing point of view, it shouldn't matter, capacity is
> > capacity. If it is a problem for node operators, they limit their
> > bandwidth in the config, problem solved. If they don't, then they just
> > run at 100% CPU, and Tor should still properly report their observed
> > bandwidth rate (unless they lie, but again, that is a another,
> > orthogonal matter).
> 
> Well, the problem is that Tor in fact doesn't properly report capacity at
> the extremes. Our measure of capacity is the most you've seen yourself
> burst in the past day -- it pretty much assumes that the pipe and other
> resources you have are static throughout the day.
> 
> So if you somehow managed to push a lot briefly but you're too busy to
> handle that level of traffic sustained, then you've overadvertised.

Again, this happens both with network load and CPU load. In fact, I
think the network load capacity differences are far more extreme and
far more common. On my box, Tor uses 10% CPU, but is constantly
prioritized below regular traffic, which is bursty and very irregular.
My node is also below the 1.5Mbyte/sec barrier, so it does not help
even out this problem for me, nor most others who share Tor with
normal traffic.

> Putting a cap on advertised bandwidth when load balancing is a crude way
> to account for this. Making our bandwidth estimate more complex may also
> work, but then we have to figure out what's better. :)

If it has to be fixed one place, it is probably best done by not
advertising burst traffic. That problem is independent of this patch,
though, IMO.

-- 
Mike Perry
Mad Computer Scientist
fscked.org evil labs
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
URL: <http://lists.torproject.org/pipermail/tor-dev/attachments/20070718/9134bb83/attachment.pgp>