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

[Roger Dingledine <arma@xxxxxxx>]

On Thu, Apr 22, 2010 at 11:31:45AM +0200, Olaf Selke wrote:
> Roger Dingledine wrote:
> > 
> > Attached is that patch. It only applies against git head, so if you
> > prefer tarballs, check out
> > http://freehaven.net/~arma/tor-0.2.2.12-alpha-dev.tar.gz
> 
> at least the situation has changed :-)
> 
> On my tiny middleman "Ilka" the number of open tcp sessions dropped from
> 10k to less than 500 and so far not going to increase again. Good! This
> should solve most router problems for residential users.

Great.

> The situation on my exit "blutmagie" is somewhat different. Number of
> open tcp sessions dropped as well.

Great. I would expect blutmagie to still have several tens of thousands
of connections open at any given time -- my theory is that many of these
are open exit connections, which the other relays here don't have.

> On the other hand ratio of new
> connections per second increased. So far I'm not sure how throughput/cpu
> load ratio is affected. Roger's patch has been started this morning at
> 07:30 local time. Pls look here http://torstatus.blutmagie.de/public_mrtg/

Interesting. One option is that you're hanging up on them and they're
coming back soon after, causing more connections per second. Another
option is that you're now able to handle more connections than before. Are
the new connections we're seeing only incoming connections, or do they
also count outgoing connections? Maybe the lower ram load is allowing
more exit connections to actually succeed. :)

> By now I no longer consider the cpu load issue being a bug. It looks
> like even a fast Intel E8600 cpu can't cope with blutmagie's exit
> traffic. Investing in new hardware won't help either as long as nearly
> all cpu time is spent within the main thread.

An oprofile of your Tor relay would be useful at some point. I bet we're
doing reasonable things that turn out to be really unreasonable when
you have a hundred thousand things in the list we loop over. :)

--Roger