[Codel] sawtooth

[Codel] sawtooth

[Simon Barber]

I've been thinking about codel's behaviour with a single TCP stream. TCP 
in steady state exhibits a sawtooth cwin behaviour. cwind rises by 1 
each RTT until the queue drops a packet, then cwind halves, and repeat.

With codel the earliest a packet will be dropped is 100ms after the 
queue size grows to the point the sojourn is 5ms. At this point the cwin 
is halved. For small RTTs or with multiple flows this may ensure good 
utilisation, but for large RTTs and a single flow surely dropping this 
early could result in starvation?

Simon

Re: [Codel] sawtooth

[Kathleen Nichols]

For a single TCP using reno and with an RTT at the nominal 100ms level,
you can expect to see about 75% utilization of the link. If you use cubic,
you do better. If you tolerate more queue, you get more utilization but
more delay. Van talked about how we pick the set points at the IETF,
http://www.ietf.org/proceedings/84/slides/slides-84-tsvarea-4.pdf

For more on this, you can see our old Red in a Different Light draft,
page 6,
http://www.cnaf.infn.it/~ferrari/papers/ispn/red_light_9_30.pdf

It's important to figure out what your objective is and measure that.
There is
the link utilization for a single TCP flow, the goodput of a particular
flow,
the delay experienced. Then there is peformance of more than one flow.

	Kathie

On 3/27/13 4:34 PM, Simon Barber wrote:
> I've been thinking about codel's behaviour with a single TCP stream. TCP
> in steady state exhibits a sawtooth cwin behaviour. cwind rises by 1
> each RTT until the queue drops a packet, then cwind halves, and repeat.
> 
> With codel the earliest a packet will be dropped is 100ms after the
> queue size grows to the point the sojourn is 5ms. At this point the cwin
> is halved. For small RTTs or with multiple flows this may ensure good
> utilisation, but for large RTTs and a single flow surely dropping this
> early could result in starvation?
> 
> Simon
> _______________________________________________
> Codel mailing list
> Codel@lists.bufferbloat.net
> https://lists.bufferbloat.net/listinfo/codel