Re: [Dyncast] CAN BoF issues and the next steps

Hi Jens,

Yes video cares more about storage, while in some cases the storage could also be seen in the scope of compute, if we consider this as a factor from the service instance side. More request will cause a heavy load.

Though there are some existing methods to guarantee the video's experience, if we want, I think video service could also benefit from CAN to get the video content dynamiclly.

Regards,
Peng

liupengyjy@chinamobile.com

From: Jens Finkhaeuser
Date: 2022-04-14 15:50
To: Linda Dunbar
CC: tony.li@tony.li; wangaijun@tsinghua.org.cn; dyncast@ietf.org; liupengyjy@chinamobile.com; luigi.iannone=40huawei.com@dmarc.ietf.org
Subject: RE: [Dyncast] CAN BoF issues and the next steps
Linda,

yes, User Application or User Agent.

This is describing how the system worked in the past. It is not strictly required that a system should work like this in future.

However, it is a fairly special edge case: video streaming doesn't so much care about "compute" as it cares about storage; this distinction gets a bit blurred when a video stream is produced on-the-fly/live.

Keeping this distinction in mind, there are hard timeouts past which a UA no longer cares about a (live) video frame. One way to deal with this is to adopt solutions similar to time sensitive networking where the UA provides such timeout information up front with every request - or the UA discards expired data itself. Since it's nontrivial to map a video frame to some byte sequence without decoding the byte sequence, the latter was the better choice.

I'm not sure what the intended scope of CAN is, and whether this kind of use case is of interest. But in case it is, I hope my description helps.

Jens

------- Original Message -------
On Wednesday, April 13th, 2022 at 20:24, Linda Dunbar <linda.dunbar@futurewei.com> wrote:

Jen, 

Does the “UA” mean “User Application” ? 

So your solution requires the applications on the UE  to query the “Central Repository”? 
When a UE roam from one Cell tower to another in the middle of the communication, does your solution require the Application on the UE to periodically query the “Central Repository” to select the optimal path to the “Source”? 

Thank you
Linda

From: Dyncast <dyncast-bounces@ietf.org> On Behalf Of Jens Finkhaeuser
Sent: Wednesday, April 13, 2022 4:07 AM
To: tony.li@tony.li; wangaijun@tsinghua.org.cn
Cc: dyncast@ietf.org; liupengyjy@chinamobile.com; luigi.iannone=40huawei.com@dmarc.ietf.org
Subject: Re: [Dyncast] CAN BoF issues and the next steps

Hi all,

speaking from practical experience with the edge case of streaming video from the "best" source, it may be desirable for the UA to make the decision locally, as it's also best suited to deciding if a response no longer matters.

We'd query a central repository for likely sources with some compute aware ordering, and then use path aware ordering at the UA for narrowing down the candidate list.

In our use case, we would query all sources, update their ordering on actual RTT, and send ACK equivalents to all, independent of which source provided each packet.

This list then periodically got refreshed.

This could conceivably be done by the ingress, use unicast for all sources, etc. but it would need some transparency. That is, E2EE would likely place this at the UA rather than an ingress router.

The point here is probably that it's likely best to differentiate between where the path aware and the compute aware metrics get aggregated and evaluated, at least for some types of application. 

Hope that makes sense,
Jens

-------- Original Message --------
On 13 Apr 2022, 05:27, Tony Li < tony.li@tony.li> wrote:

Hi Aijun,
My understanding of the requirements was that a particular UE was supposed to be bound to a server for the lifetime of a ’transaction’ and that includes across the UE rehoming to a different source. Thus, the entire network needs to make a consistent and fixed decision per UE. Doing so in a dynamic environment would seem to be most challenging: you would need to synchronize forwarding state across the entire network instantly.
Making a single decision at the ingress and propagating that state seems somewhat easier.
And easier still is Joel’s proposal: have the UE pick one (from a centralized broker?) and then rely on unicast. :-)
Regards,
T
> On Apr 12, 2022, at 7:59 PM, Aijun Wang <wangaijun@tsinghua.org.cn> wrote:
>
> Hi, Joel:
> If you use binding address behind the ANYCAST address, it is possible. But if you use the ANYCAST address directly, you can't.
> For my understanding, the latter scenario is more popular.
>
>
> Best Regards
>
> Aijun Wang
> China Telecom
>
> -----Original Message-----
> From: Joel M. Halpern <jmh@joelhalpern.com>
> Sent: Wednesday, April 13, 2022 10:55 AM
> To: Aijun Wang <wangaijun@tsinghua.org.cn>; 'Luigi IANNONE' <luigi.iannone=40huawei.com@dmarc.ietf.org>; liupengyjy@chinamobile.com; 'dyncast' <dyncast@ietf.org>
> Subject: Re: [Dyncast] CAN BoF issues and the next steps
>
> If the ingress edge does the calculation, makes the determination, and tunnels the traffic to the right place then the underlay routing system does not need to know anything about these metrics or the decision processes made by the edge.
>
> Yours,
> Joel
>
> On 4/12/2022 10:52 PM, Aijun Wang wrote:
>> Hi, Luigi:
>> Why only the ingress need such decision? I think all the routers
>> in-path need such information(routing metric +compute metric), to
>> achieve the optimal "instance selection".
>>
>> Best Regards
>>
>> Aijun Wang
>> China Telecom
>>
>> -----Original Message-----
>> From: dyncast-bounces@ietf.org <dyncast-bounces@ietf.org> On Behalf Of
>> Luigi IANNONE
>> Sent: Tuesday, April 12, 2022 3:04 PM
>> To: Aijun Wang <wangaijun@tsinghua.org.cn>; 'Joel M. Halpern'
>> <jmh@joelhalpern.com>; liupengyjy@chinamobile.com; 'dyncast'
>> <dyncast@ietf.org>
>> Subject: Re: [Dyncast] CAN BoF issues and the next steps
>>
>> Hi,
>>
>>> But, with the placement of the ANYCAST application servers closing to
>>> the users in different sites, the bottleneck to influence the E2E
>>> application performance is not only the network metric, the metric
>>> for the application servers play a major role now.
>>> It is time to consider both the network metric and application server
>>> metric together to achieve such goals.
>>
>> I think that Joel is not against the above (routing metric +compute
>> metric = instance selection).
>> I think that he is more inline with Dirk's position, meaning that it
>> is not necessarily the routing layer that has to be "enhanced" with
>> compute metrics.
>> Rather, an in-path decision based on both metrics should be made by
>> some (CAN ) element.
>> My personal take is that the ingress is well suited for that (since
>> for sure it is in-path).
>> Then you have the choice of various ways on how to steer the traffic.
>>
>> Ciao
>>
>> L.
>>
>>
>>
>>
>> --
>> Dyncast mailing list
>> Dyncast@ietf.org
>> https://www.ietf.org/mailman/listinfo/dyncast
>>
>
> --
> Dyncast mailing list
> Dyncast@ietf.org
> https://www.ietf.org/mailman/listinfo/dyncast
--
Dyncast mailing list
Dyncast@ietf.org
https://www.ietf.org/mailman/listinfo/dyncast