Re: [tcpm] flow control and fast recovery

["Scheffenegger, Richard" <[e-mail]>]

HI Yoshifumi,

> > Note once the pipe is below ssthresh PRR will do slow-start, just like
> > normal CC.
> 
> Oh. I see. Thanks for pointing out.
> Then, I think at least we should avoid slow-start increase when TCP
> retransmits retransmited packets, although I don't have a clear idea for
> the best behavior in this kind of situation, yet.
> Because this is basically an indication of packet loss during recovery
> phase.


So far, lost retransmission detection, or fast recovery of a lost retransmission, has not been mentioned in any RFC; the default behavior of stack compliant with the RFCs is to run into a RTO - which is a pretty hefty reaction to continued loss (congestion indication) during recovery. OTOH, one particular OS has demonstrated, that by simply ignoring the congestion signal due to lost retransmissions also doesn't lead to an instant network meltdown.

Once I've finished what I'm currently focused on, I would like to submit a paper to sketch out lost retransmission detection and recovery (with SACK). I think such a draft would be the proper place to mention the expected behavior.

Best regards,

Richard Scheffenegger