Re: constrained socket buffers patch

[Roger Dingledine <arma@xxxxxxx>]

On Fri, Jul 13, 2007 at 03:44:49AM +0200, Juliusz Chroboczek wrote:
> I'd like to know how this will interact with tor's circuit selection.

Good question. The first answer is "it won't change it at all",
but you're right, it may be more complex than that.

> If I understand this patch correctly, it sets the SO_SNDBUF and
> SO_RCVBUF socket options so as to limit the maximum size of the TCP
> send and receive windows respectively.
> 
> Now the throughput (``bandwidth'') of a TCP connection is limited by
> window/rtt.  What this means is that with ConstrainedSockets enabled,
> your tor server will have basically unlimited throughput on a local
> connection, but be limited to roughly 40 kB/s per connection (that's
> bytes, not bits) over a transatlantic link[1].
> 
> Perhaps one of the tor developers could clarify whether tor's circuit
> selection and congestion control can deal with server's whose per-
> connection throughpout (but not necessarily aggregate throughput) is
> widely dependent on where you come from?

I'm interested to see your [1] where you do the numbers. :)

Tor servers advertise aggregate throughput. In fact, it's often the
case that some link or another can't push many bytes, either because
it's slow or because that node is rate limiting a lot already.

So looking at it another way, there's enough variety in servers already
that it's not clear this would add many new problems to the network --
it already has lots of problems.

--Roger