[e2e] Congestion control as a hot topic in IETF
Fred Baker (fred)
fred at cisco.com
Sat Mar 9 05:31:23 PST 2013
On Mar 4, 2013, at 5:07 PM, "Scharf, Michael (Michael)" <michael.scharf at alcatel-lucent.com> wrote:
> Some time ago, I wondered myself whether an Internet without congestion control could work, or not. My (somehow short) insight can be found on page 67 of: http://www.ikr.uni-stuttgart.de/Content/Publications/Archive/Sf_Diss_40112.pdf
Well, we have experience with networks using Internet technology with inadequate congestion control. That's a good description to the Internet before TCP congestion control was deployed, and I can think of some other cases in which people learned about it the hard way in private networks.
The question, I suppose, is 'how "controlled" is "controlled well enough"?'. There is work right now on deterministic (read: TDM-like) networking, including a project funded through FIRE a few years back, and commercial product related to avionics and industrial automation. I personally think it's a little oversold, but you should know that the idea is out there. In a network that is deterministic end to end, one could argue that the only congestion control required is for a host to not congest its own queues, as no other queue would (by design) become congested or for that matter introduce more than the design quantum of delay.
I'm personally of the opinion that loss-based congestion controls are inefficient, for the simple reason that a mechanism that tunes to the knee will generally achieve the same bit rate end to end as one that tunes to the cliff (they will both, given that they have enough data to do so, fill the bottleneck); However, one that tunes to the cliff has to assume and account for some amount of self-inflicted loss, which implies recovery delays and a reduction in end to end bit rate comparable to the retransmission rate. CAIA CDG has, I think, characteristics that would enable it to eliminate a number of data center issues and issues on the wide wooly Internet.
Yes, we need some form of congestion control. We have proven that to ourselves time and again. The question is "with what characteristics, seeking to achieve what goal?"
More information about the end2end-interest