On Thu, Apr 18, 2013 at 5:08 PM, vmonmoonshine@gmail.com wrote:
Zack Weinberg zackw@panix.com writes:
and the retransmit of packet 1 ought to be happening on a *different* connection, if we have one. (Which steg are you using?)
I'm using nosteg steg. I thought if something is going to work, better, it works with the simplest steg. nosteg only open one connection and pass everything through it as long as there is no reason to drop it, for bad header etc.
OK. We shouldn't even try to retransmit with a 1-connection steg mode (unless it's not using TCP ... worry about that later)
So I'm not understanding exactly what the "dropper proxy" does. Does it prevent TCP from providing reliable delivery? If so, how?
It sits at socket_read_cb and sometimes doesn't copy what libevent's has read from one side of communication, into the buffer of the other side.
Ah. Yeah, that's not going to work. TCP will think that the data _has_ been delivered, so the lower-level retransmit that we're relying on will never happen.
So, what you say, means that libevents socket_read_cb calls are more refined than an entire TCP packet that TCP guaranteed to deliver. I.e, libevent break something that TCP guaranteed to deliver in smaller parts.
TCP is a stream-oriented protocol. It guarantees to provide reliable, ordered delivery of a _sequence of bytes_. It does _not_ guarantee anything whatsoever about packet boundaries. In particular, the amount of data libevent hands you in one read callback is completely meaningless.
If this is true, I guess because TCP is giving us a stream there is no way for us to know where to drop to have a legitimate simulation of real life packet drop in that stream, unless I incorporate part of chop in the dropper proxy to read their headers and detect end of packet.
What you need to do is implement the dropping _below_ TCP, so that TCP is aware of it and does do its retransmits. You can do this on Linux with netem http://www.linuxfoundation.org/collaborate/workgroups/networking/netem and on many of the *BSDs (including OSX) with dummynet http://info.iet.unipi.it/~luigi/dummynet/.
ST absolutely *does* need a congestion control mechanism, though, to prevent the entire circuit from getting killed because it overran the fixed-size reassembly queue, and as long as we have to do that ...
Then maybe I just start an axe timer whenever when I want to send and the transmit queue is full and delete the timer whenever the queue moves.
How do you know that your peer's receive queue is full, if it doesn't tell you?
But we probably could get away with something much simpler if we don't bother doing retransmits, e.g. something akin to Tor's SENDME cells.
zw