[e2e] Link Aggregation

[Valentin Ossman]

Dear Spencer,

Thanks for the analysis. 
You are completely right. 
A simple solution can be sending ACK only for packets received on one
port (acknowledging also packets sent on the second port). Having ACKs
only on one path will make life much simpler because there will not be
out-of-order ACKs to trigger retransmit (am I right here?).

For example:

Packet sequence
Sender                     Receiver
Port 1: ---*3*-----*1*---> :Port 1
Port 2: -----*4*--*2*----> :Port 2
ACK sequence (only port 2 sends ACK)
Port 1: <------IDLE------- :Port 1
Port 2: <----*2*----*4*--- :Port 2

ACK *2* was triggered by packet *2* on port 2 and ACK *4* was triggered
by packet *3* on port 1 (packet 4 already arrived).

Also, to ease reassembly, maybe RDMA may be used over TCP.

Valentin

> Dear Valentin,
> 
> The problem would be that once the "last router" starts putting
> packets from a single flow on two different ports, we KNOW that
> the packets will not be routed along the same path, and there's
> no "resequencer" (short of the destination host) that will
> ensure that we're not generating spurious Fast Retransmits at
> the destination host.
> 
> Packets that are out-of-sequence by three packets aren't that
> hard to generate, if you really have two gigabit streams being
> independently routed through the network.
> 
> There are limited environments where what you're talking about
> might work (3GPP uses a sequenced tunneling protocol, for
> instance), but the general case is harder.
> 
> In the scenario you describe, if you spend much time in Fast
> Recovery, you quickly lose the advantage of the second link
> (because each Fast Recovery reduces cwnd by 50%).
> 
> TSVWG has active proposals for a "response to spurious Fast
> Retransmit detection" (see
>
http://www.ietf.org/internet-drafts/draft-ietf-tsvwg-tcp-eifel-response-
> 03.txt),
> but I believe the proposals for adjusting the threshold for Fast
> Retransmit are still floating around the research community.
> 
> I hope this is useful to you,
> 
> Spencer Dawkins
> 
> --- Valentin Ossman <valentin at tehutinetworks.com> wrote:
> > Or, even better, let's suppose that the last router (connected
> > to the
> > station network) knows about this link aggregation and is
> > distributing
> > the packets in-order and evenly between the 2 ports to achieve
> > higher
> > bandwidth.
> >
> > Valentin