[Ndn-interest] Hop-by-Hop Flow Balance

[Klaus Schneider]

Dear Dr. Solis,

I agree that there are some missing pieces to the Hop-by-Hop flow 
balance. Can you give us some pointers to the "other ways to do 
flow-balance and flow-control" ?

Can they solve the problem that the data packet size is essentially 
unknown to routers and can be huge?

Best regards,
Klaus Schneider


> Message: 2
> Date: Mon, 14 Mar 2016 23:09:34 +0000
> From: <Ignacio.Solis at parc.com>
> To: <aa at CS.UCLA.EDU>, <ndn-interest at lists.cs.ucla.edu>
> Subject: [Ndn-interest] Hop-by-Hop Flow Balance
> Message-ID: <F7A301F8-CE16-42EC-855E-CB8CB1472A00 at parc.com>
> Content-Type: text/plain; charset="utf-8"
>
> [ Disclaimer: CCN currently uses flow balance as well ]
>
> The current Hop-by-Hop Flow Balance is nonsense.
>
>
>
>
>
> On 3/10/16, 11:46 PM, "Ndn-interest on behalf of Alex Afanasyev" <ndn-interest-bounces at lists.cs.ucla.edu on behalf of aa at CS.UCLA.EDU> wrote:
>> [6] **Hop-by-Hop Flow Balance**:
>>    Over each link, one interest packet should bring back no more than one data packet.
>
> Seriously, who thinks this actually works?
>
> Let me quote the webpage ( http://named-data.net/project/ndn-design-principles/ ):
> "[6] Hop-by-Hop Flow Balancing: Over each link, one interest packet should bring back no more than one data packet.
> Hop-by-hop flow balancing enables each node to control load over its links. By deciding to sending interest over a link, router commits bandwidth for the returned data. By limiting the number of interests sent, each router and client node in the network control how much data it will receive.
> "
>
> Either there is a lot of information missing here to justify why this is so, or this is very na?ve.
>
> First, if what you want to do is limit the number of content objects (or packets) returned, you don?t need to send one interest.  _Specially_ for NDN, which has prefix matching, you could send one interest with a count number (10) and expect to receive 10 content objects back.  There is no reason why I need to send 10 exact copies of the same interest.   Even if the interests had small variations, why send 10? Why not send 1 + the 10 deltas?   I guess it?s possible you may call that part of the ?network adaptation layer?, who knows.
>
> Also by requiring 1-to-1, you are always requiring an overhead (on the requester side) that is quite high. If you think of today?s type of networks, where a packet (internet sized) is around 1500 bytes, that means that even if we send interests of 30 bytes, we are incurring quite a bit of overhead in the upstream. This becomes considerable when doing high bandwidth video.
>
> Please explain why the 1-to-1 is good.
>
> Second, NDN allows very large packet sizes.  So, when I send 1 interest, I don?t now if what I?m going to get back is 1 byte or 18 exabytes.  How do routers use this information to control how much data they?re going to receive? Are they going to reserve 18 exabytes of traffic time?
>
> If this principle were to be re-written as:
> ?Allow network nodes to participate in flow control? then the actual engineering solution might be able to achieve this.
>
> Finally, at least we should acknowledge the limitations this type of approach requires; like symmetrical forwarding.
>
> It would be awkward if the only way for NDN to work over Satellite links would be to break the principles.
>
> Nacho
>
>
> PS. Yes, there are people in this community who have looked at other ways to do flow-balance and flow-control. Maybe we should be learning from those and not just claiming as principle what we do today because we don?t want it questioned.
>
> --
> Nacho (Ignacio) Solis
> Protocol Architect
> Principal Scientist
> Palo Alto Research Center (PARC)
> +1(650)812-4458
> Ignacio.Solis at parc.com
>