[e2e] TCP ex Machina

[Keith Winstein]

Hello,

I'd be grateful for the list's feedback re: our new paper on
computer-generated end-to-end congestion-control schemes.

The paper, FAQ, code, and steps for reproducing the results are up at:
http://web.mit.edu/remy

We find, essentially, that computers can create better congestion-control
protocols than the best human-designed schemes to date -- if given a clear
statement of the prior assumptions that the protocol is entitled to have
about the network going in, and a well-stated objective. (Which could be
throughput fairness, throughput over delay, average flow completion time...)

Our program, Remy, has generated end-to-end congestion-control schemes that
get more throughput, less queueing delay, and more fairness than TCP Cubic,
Compound TCP, TCP Vegas, etc. Relevant to this list, we were surprised to
find that these computer-generated schemes (running over "dumb" DropTail
gateways in simulation) can even outperform schemes like XCP and
Cubic-over-sfqCoDel, which depend on intelligence inside the network, and
in XCP's case depend on prior knowledge of the exact wire speed at the
gateway. It seems like an unexpected victory for end-to-end algorithms that
we're working to explain.

To be clear, the work is academic research, in ns-2, and has a long ways to
go (including a real implementation, and understanding HOW these crazy
computer-generated schemes work) before we would think anybody would want
to deploy it in a datacenter under the control of one entity, much less
over the Internet

But in our view, making the transport layer be simply a function of (a)
what the lower layers are doing and (b) what the users and applications
want is a way to get a more evolvable Internet, and avoid the problem where
network infrastructure does unfortunate things to try to mask loss and
out-of-order delivery from TCP.

We are presenting this next month at SIGCOMM so will be grateful for any
early feedback the list can supply.

Regards, and thank you,
Keith