Roger Dingledine wrote:
On Mon, Jan 23, 2006 at 12:36:53PM +1000, glymr wrote:yes, that's not a good assumption to make however. average uptime is a more useful metric, when a system has been up for a long time it may be just about to go down. also, for irc users, a connection which can stay up for 8 hours or more is regarded as quite adequate by most.
These rules could be used to weight classes of ports, a node could keep a history of its uptime, and report its average uptime value accumulated over time to the directory. This would help for choosing interactive traffic routes, the longer the average uptime, the greater the chance of it being picked on interactive circuits.
Actually, we already do something like this. Nodes report their uptime, and we assume that a long uptime implies that it will stay up.
what i mean is, different ports, for different protocols, have different traffic characteristics. irc, silc, msn, aim/icq, jabber etc, these are all long lived and mostly low bandwidth connections, whereas web browsing is short lived and bursty. they should be treated differently.Tor could automatically select it's preference for the different traffic classes according to these values. At this point, without an automated system to do this, it can be done by users (as I am doing) - by using a rate-limiting system (netlimiter) and allowing only a small set of interactive traffic types through (in my case, irc and silc) - since tor precludes the use of file transfers on these two protocols, I set the rate limiting between 2 and 4kb/s depending on whether I am downloading more or chatting more.
I'm not sure I understand this part. You are using an external tool to
traffic shape the stuff you send into Tor? Or you are using an external
tool to traffic shape the stuff that exits from your Tor server?
I think that what i have described is the, at this point at least, best answer to the problem of improving long lived interactive connections.However, I think it would be a worthwhile addition to the system by which Tor does its routing to use these rules in both the production of an uptime and bandwidth average, which is used by clients to select a pair of different circuit classes, interactive and high volume. High volume traffic usually is short lived, and interactive traffic is usually long lived. By specialising the circuits according to these rules one would find that interactivity is better promoted, and separated from volume.
Right, we do this too.
Check out the man page entry:
LongLivedPorts PORTS A list of ports for services that tend to have long-running connec- tions (e.g. chat and interactive shells). Circuits for streams that use these ports will contain only high-uptime nodes, to reduce the chance that a node will go down before the stream is finished. (Default: 21, 22, 706, 1863, 5050, 5190, 5222, 5223, 6667, 8300, 8888)
In 0.1.0.x, "high-uptime" is defined as "claiming an uptime of at least 24 hours." In 0.1.1.11-alpha, high-uptime is defined as meeting or exceeding the median uptime of running valid servers. On the current Tor network this tends to vary between 2 to 3 days.
But this clearly does not totally solve the problem: long-term connections over Tor do still break. Part of this is because the Tor network is very young and still quite dynamic, so people are upgrading, restarting their servers, and so on. There is also clearly a tension between scaling the network (and thus being more flexible about the volunteers we can make use of) and providing stable links.
Another nice approach would be to be able to move streams to a new circuit if the current one dies. But this has its own problems: http://wiki.noreply.org/noreply/TheOnionRouter/TorFAQ#MigrateStreams
What other approaches are there that might work?
--Roger
David