[xcp] Some remarks on XCP

[Michael Savoric]

Yuri Pryadkin wrote:
> 
> On Thursday 11 November 2004 02:14 am, Michael Savoric wrote:
> > Yuri Pryadkin wrote:
> > > Would you please explain how you propose to change XCP to hand out
> > > absolute feedback?
> >
> > It's very simple. Avoid dividing the positive and negative feedback
> > by H_CWND in the router (I refer to the earlier XCP papers).
> > Send this absolute feedback to the sender. The sender changes its CWND
> > to the feedback.
> 
> If you refer to [1], the pseudo-code there in Appendix A is, I quote:
>   On packet departure do:
>         pos_fb = xi_p * H_rtt * H_rtt / H_cwnd
>         neg_fb = xi_n * H_rtt
> Are you suggesting to remove division from the first quoted line and
> add a multiplication to the second?  I'd like to be sure we are
> talking about the same thing.
> ---------
> [1] Dina Katabi, Mark Handley, and Charles Rohrs, "Internet
> Congestion Control for High Bandwidth-Delay Product Networks."
> ACM Sigcomm 2002, Pittsburgh, August, 2002.
>

Yes, that's what I propose.

> In the later incarnations of XCP that have packet sizes factored in
> and where the switch has been made from congestion window to
> throughputs, if I understand your proposal correctly, it boils down
> to:
>         1.  Calculate incremental change in throughput as per XCP
>             (delta_throughput)
>         2.  expected_packets_per_ci = old_throughput * CI / packet_size
>         3.  new_throughput = old_throughput
>                 + delta_throughput*expected_packets_per_ci
> (CI stands for control interval, and Cp and Cn are xi_p and xi_n
> adjusted for throughput).
> 
> Router per packet calculations do get simpler.  Because
> delta_throughput is calculated as
>         Cp * packet_size/throughput - Cn * packet_size
> after multiplying by throughput * CI / packet_size you end up with
>         new_throughput = old_throughput
>                 + (Cp * CI - Cn * CI * old_throughput)
> so there is no division in the per-packet processing.  This is a big
> advantage.  Another advantage is that round-off errors do not
> accumulate at the router if you're doing integer arithmetics.
> 
> However, there may be some concerns about the stability of this
> method.  Especially when flows have widely different RTTs and/or
> traffic is bursty, oscillations may be greater.
> 
>   -Yuri

Reducing the per-packet costs in the router is only a positive
side-effect of my proposal. More important is that negative
feedback is sent much faster back to the sender.

Best regards,
Michael
 
======================================================
E-Mail:		savoric at ee.tu-berlin.de

Phone:		(+49 30) 314-29244

Fax:		(+49 30) 314-23818

Postal address:	Technical University Berlin
		Telecommunication Networks Group (TKN)
		Einsteinufer 25, 10587 Berlin
======================================================