Re: [arch-d] [Network-tokens] [Apn] Questions for APN: Q#5

[Yiannis Yiakoumis <[e-mail]>]

Hi Lars,

On Wed, Sep 30, 2020 at 12:28 AM, Lars Eggert < [e-mail] > wrote:

> 
> 
> 
> Hi,
> 
> 
> 
> 
> On 2020-9-26, at 2:22, Yiannis Yiakoumis < yiannis@ selfienetworks. com (
> [e-mail] ) > wrote:
> 
> 
> 
>> 
>> 
>> One of the first successful proof-points for SDN was Google claiming that
>> they moved their WAN utilization to ~100%, instead of over-provisioning.
>> It saved them lots of money. Two of the design principles that led them
>> there were
>> i) moving from "all packets are equally important" to "allocate resources
>> based on application-level priorities" and ii) moving from "TCP flows
>> regulated by "fair share" mechanisms" to "measure demands, and shape flows
>> at the endpoints".
>> 
>> 
> 
> 
> 
> I think these claims were made for Google's private B4 network, right?
> AFAIK that is a *very* different network from the public Internet. From
> what I recall, each transfer over B4 is scheduled over the course of a day
> based on various flow properties that are required to be specified,
> following some giant minmax-like computation.
> 
> 
> 
> 

Yes, this is for B4. While the network is different, the idea that you can build a much more cost/performance efficient WAN still stands.

> 
> 
>> 
>> 
>> I can imagine offering low-latency SLA for 1% of the traffic (say video
>> calls), is much cheaper and easier than doing this for all Internet
>> traffic (with the cost varying whether you are in Europe/US or an emerging
>> market, rural or urban, a WISP, a low-orbit satellite network, etc). Could
>> anyone from operator-land comment on the economics of
>> oversubscription/overprovision/SLAs in the mobile world?
>> 
>> 
>> 
> 
> 
> 
> Sure. But that could be done with existing mechanisms, like DSCP marking.
> Remarking/bleaching does happen as traffic gets routed between networks,
> but since the stated APN use case here is within an operator network, that
> is under that operator's control.
> 
> 
> 
> 

DSCP marking works within a single administrative domain where you have full control of endpoints and the network. The moment you want to incorporate end user devices and or third-party servers that sit out of the network, it doesn't work well. My comments are not per APN usecases per se, but rather the bigger traffic differentiation picture. (if I understand correctly, APN tries to limit things within an administrative domain to avoid having to deal with trust, security, neutrality, etc).

> 
> 
>> 
>> 
>> (btw: we already have SLAs today --- VoLTE is IP traffic routed through
>> dedicated radio and wired resources and has materially better network
>> quality and reliability than plain VoIP, and when you sign-up for a
>> broadband connection you can get 10Mbps/100Mbps/1Gbps etc).
>> 
>> 
>> 
> 
> 
> 
> Yup. And that happens without APN in the picture.
> 
> 
> 
> 

This happens by having vertical integration for a specific use case (phone app, phone driver, network control plane, network data plane ). I would really like to get VoLTE's reliability into my Zoom/FaceTime/WhatsApp calls, but it is not possible.

> 
> 
>> 
>> 
>> If over-provisioning wins (in 10 years) it will likely be the outcome of
>> the network community's failure to find a good solution for traffic
>> differentiation, rather than a verdict that these use cases were not
>> important enough.
>> 
>> 
>> 
> 
> 
> 
> The problem isn't that we don't have good solutions for traffic
> differentiation. We have tons. The problem is that none of the existing
> ones have seen much deployment - but not because they don't work, because
> of other issues. I don't see APN to offer materially different/better
> deployment incentives than what we already have, and so IMO it would share
> the same fate.
> 
> 
> 
> 

I would disagree here. We do have several traffic differentiation deployments that use DPI and application signatures to map traffic into SLAs at a fine granularity. It is largely inefficient cost and performance-wise, error-prone, and violates privacy and neutrality in certain occasions. But that's the one solution mostly used, for the lack of better alternatives.

Best,

Yiannis

> 
> 
> 
> Thanks,
> Lars
> 
> 
>> 
>> 
>> FWIW: there are many issues to debate around fine-grained traffic
>> differentiation beyond overprovisioning (like security, scalability,
>> privacy, net neutrality), and there are different approaches one could
>> take (APN and network tokens have different design principles and
>> priorities on these trade-offs).
>> 
>> 
>> 
> 
> 
>