Re: [Tsvwg] Re: HighSpeed TCP for Large Congestion Windows

So, lets assume a cwnd of 800 packets and that 80 are lost.

> I'm not an expert in SACK, but if I'm not wrong, there are a
> couple of issues that can limit SACK behaviour:
>
> - no more that 3 SACK block per ACK can be accomodated

That's right (assuming timestamps, which you'd need for a cwnd of
800 packets (of much size, anyway).  But, we're going to have 720
ACKs to piggyback SACK information on.  And, 80 losses can cause at
most roughly 80 SACK blocks.  So, it seems like there are plenty of
chances for returning all the SACK information to me.

> - the pipe mechanism will limit the number of retransmission per RTT

We will be able to send 400 packets in the RTT following loss
detection.  That is five tims what we lost.  I think that should
suffice to allow for the retransmission of everything.

> - SACK will fall back to RTO in case a retransmitted packet is lost

Not necessarily.  If you look at our current draft
(draft-allman-tcp-sack-12.txt) then yes, if a retransmit is dropped
we will RTO.  And, there is probably a higher chance of losing a
retransmit when the window is larger.  But, you'd have to do the
experiment before I believed this was a huge problem.  And, further,
SACK *can* detect lost retransmits and retransmit again without an
RTO if the TCP sender is willing to keep more state.  I know some
folks have hashed out schemes to do this.

My high order point here is that I do not think there is anything
fundementally wrong with SACK such that it will not work with large
cwnds in a high speed environment.  I believe it scales.

allman

--
Mark Allman -- BBN/NASA GRC -- http://roland.grc.nasa.gov/~mallman/
_______________________________________________
tsvwg mailing list
tsvwg@ietf.org
https://www1.ietf.org/mailman/listinfo/tsvwg