Re: Tor client performance (was Re: URGENT: patch needed ASAP for authority bug)

[Sebastian Hahn <mail@xxxxxxxxxxxxxxxxx>]

On Apr 21, 2010, at 3:13 PM, Olaf Selke wrote:
> output from the first 20 minutes after startup can be found here:
> http://selke.de/txt/notices.html
>
> Flight level after 20 min tor uptime is 30k+ tcp connections.

Interesting. This matches what I see on fluxe3 (although on a
much larger scale, of course), in that around 10% of
connections seem to not have a circuit attached.

I think we should explore doing three things:
a) Make clients close connections immediately when they
close a circuit
b) Make relays close those connections the client hasn't (yet)
closed
c) Make clients close circuits used for begin-dir stuff a lot
sooner than after the 10 minutes timeout that exists now.
I don't think the chance to possibly cannibalize the circ later
justifies the load this is causing.

Sebastian