Roger Dingledine:
On Tue, Mar 25, 2014 at 01:41:43PM -0700, Mike Perry wrote:
Nicholas Hopper:
I modified Aaron Johnson's TorPS simulator to simulate 50K clients
** The reason I ask is because I suspect there is actually an interplay between the current circuit build timeout code and the pool of 3 guards. The CBT cutoff will tend to cause you to avoid whichever of your guards may be temporarily overloaded at a given time.
Does Aaron's TorPS implement the circuit-build-timeout behavior?
I wouldn't be surprised if it skipped over that, since (for example) where would you get the realistic data about network latencies and relay delays to be able to simulate it?
Oh, hrmm.. Well, it appears TorPS is just a tor controller, so I believe that CBT should still time these circuits, but if this simulator runs on a timer for extending hops on circuits (or even experiences delays while servicing control port CIRC events), it will effectively destroy how CBT behaves (and possibly alter the results in other ways).
This now makes concerned we may still not be fully evaluating the effects of this change with real Tor client behavior. I suppose we need to find out more about exactly how it builds circuits and what torrc options it uses.
But in either case, we still need a controlled experiment between 3 guards vs 1 guard at the same cutoff, otherwise we're sort of just sweeping any performance problems caused by the change under the rug.