Setting congestion window on loss detection

[hiren panchasara]

On 10/07/15 at 12:54P, hiren panchasara wrote:
> Found this issue about a month ago and started a discussion on -net:
> https://lists.freebsd.org/pipermail/freebsd-net/2015-September/043249.html
> 
> I feel this forum is a better place to discuss this further now.
> 
> Problem: We set cwnd to 1mss when we detect loss via arrivals of 3 dupacks.
> That is wrong as we severely underutilizing network capacity by doing
> so.
> 
> Next question is, what should we set cwnd to?
> 
> RFC6675 (TCP SACK) suggests following on detecting loss:
> ssthresh = cwnd = (FlightSize / 2)
> 
> RFC5681 (TCP Congestion control) suggest:
> ssthresh = max (FlightSize / 2, 2*SMSS)
> cwnd = (ssthresh + 3*SMSS)
> 
> (Here, FlightSize is bytes in flight.)
> 
> OR should we let whatever congestion control (CC) algo in control decide
> that value?

I also tried to look at what Linux does. It has PRR (Proportional Rate
Reduction) RFC 6937 (something I plan to work on after these initial
needed fixes/improvements) in place. Looking back pre-PRR code, linux
seems to be doing following:

cwnd = min(cwnd, FlightSize)

Here, cwnd in the equation is adjusted as per rate-halving
(draft-mathis-tcp-ratehalving-00) which says "the window is reduced by
sending one data segment for each two segments which are acknowledged".

(I am not very familiar with linux code so please correct me if that's
not the case.)

Basically, I think any of these approaches is better than what we have in
the tree right now.

Cheers,
Hiren
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 603 bytes
Desc: not available
URL: <http://lists.freebsd.org/pipermail/freebsd-transport/attachments/20151012/63923916/attachment.bin>