Re: [tcmtf] [tsvwg] M2M, WSNs and draft-saldana-tsvwg-tcmtf

[<[e-mail]>]

Brian,

a two parts response:

QoS interconnection isn't free of charges. The philosphy I am
aware of is "sending party pays". It's a sender incentive
then to ensure the traffic is marked as desired.

Traffic using TCP transport should use a QoS class optimised
for TCP transport, I'd suggest. A recurring issue during
standardisation and interconnection discussion with network
engineers is to avoid to name QoS classes by applications.
The point is to look at the required QoS properties. Transport
seems to be reasonably abstract.

Another issue: QoS means to avoid congestion (this is
not CoS, meaning "survive longer"). Under foreseeable
operational conditions, all QoS classes experience similar
forwarding (no drops and low delay/jitter).

That's why I also don't favour approaches starting from
an application and ending with a new required QoS class.
Where's the experience with the deployed classes
justifying this way of engineering?

The more basic part of your statement:

I agree, we don't need *another* set of classes. You, James and
I just don't agree on which are the ones we don't need. The
set I propopse is rather abstract. I didn't up to now catch a
single line telling me why this set of classes isn't useful.
Instead you say that you prefer a small set as I propose but
then you support James class/codepoint overspecification.
Too many classes, codepoints and standards around them
are part of the problem, they are not part of the solution.

To be precise: Interconnection agreements are simple regarding
the EF and the Best Effort class. They are work intensive
for AF based classes. Do you have an idea what could be
the cause?

No standards body besides IETF works with these many
codepoints and classes. Even if they propose pure IP
interconnections. But (obviously) each standard body
picks another subset and some of them interpret IETF AF
standards rather creative. Yes, DiffServ is flexible enough
to deal with that. Administration, operation and
interconnection negotiation however benefit if there's a
standard behviour, used by a majority of all customers and
by tailored adaptions used by a few.

Regards,

Ruediger


On 30/07/2012 07:53, [e-mail] wrote:
> Folks,
>
> avoiding a DSCP rewrite at domain boundarys is probably
> impossible - there are many independent DiffServ
> deployments by now.
>
> Al Morton will present some ITU work on the issue. We propose
> to apply a small standardised set of domain behaviours (also
> called QoS classes) at interconnection interfaces. The benefit
> is a domain remarks tranmitted/received traffic to/from the
> interdomain DSCPs.

That assumes that you trust the diffserv markings applied by your
peer. In the general case, that seems unlikely, which is why there
is always likely to be a classifier as well as a re-marker at the
boundary, and quite possibly a traffic conditioner too.

Agreed, there might be a subset of peering relationships at which
you would be prepared to trust your neighbour's DSCPs. I thought
that was one intent behind RFC 4594 and its possible update.
We surely don't need *another* set of recommended traffic classes.

However, at a peering connection where you don't have that level of
trust, you have to either reclassify all packets or mark everything
as best effort.

Every ISP needs to solve the interworking
> issue once only and should expect a reasonable end to end QoS.
>
> The proposal more tries to look at DiffServ router forwarding knobs:
> a queuing mechanism and active queue management. My view is, these
> can be optimised for an expected transport behaviour:
>
> UDP traffic (streaming or telephony, real time gaming) will most
> likely not be retransmitted. A short queue will do, active queue
> management by random drops isn't required (unless you'd like
> to apply PCN marking or the like).
>
> TCP traffic interacts with router congestion indications, queues
> may be a little bit deeper and active queue management may be
> deployed. Today it is RED, ECN or the like in future.

How do you help the alarming amount of traffic that is streaming over
TCP?

    Brian

>
> If one looks at it from that angle, the number of useful
> QoS classes is small.
>
> Take the transport indication (TCP/UDP) as "dominating" in a
> class, not a pureness requirement.
>
> Regards
>
> Ruediger
>
>
>
> -----Original Message-----
> From: [e-mail] [mailto:[e-mail]] On Behalf Of Dan Wing
> Sent: Friday, July 27, 2012 6:27 PM
> To: 'Jose Saldana'
> Cc: [e-mail]; [e-mail]
> Subject: Re: [tsvwg] [tcmtf] M2M, WSNs and draft-saldana-tsvwg-tcmtf
>
>> -----Original Message-----
>> From: Jose Saldana [mailto:[e-mail]]
>> Sent: Friday, July 27, 2012 4:14 AM
>> To: 'Dan Wing'
>> Cc: [e-mail]; [e-mail]
>> Subject: RE: [tcmtf] [tsvwg] M2M, WSNs and draft-saldana-tsvwg-tcmtf
>>
>> Well, perhaps a solution is the use of Diffserv bits: in the draft for
>> updating rfc4594
>> (https://datatracker.ietf.org/doc/draft-polk-tsvwg-rfc4594-update)
>> there are
>> classes for real-time services and for best-effort ones. I have always
>> believed that the update of rfc4594 can be beneficial for TCMTF: you
>> can
>> build a tunnel with flows belonging to the same traffic category.
>>
>> But I don't know if the best-effort class means exactly the same as
>> delay-tolerant.
>>
>> Dan, what do you think?
>
> I agree that DSCP is the best tool we have to combine packets
> with similar characteristics.
>
> A recurring problem with DSCP is they are re-written to different
> values because a router was configured to rewrite them all the
> time or because a queue of that traffic class had built.  That
> rewriting needs to be considered (or ignored) when grouping packets
> that are 'similar'.  IMO, the fact that a packet was in a queue
> a few router hops away (and now has a different DSCP than
> its friends) doesn't change that it should be grouped with all
> of its friends into a multiplexing tunnel.
>
> -d
>
>
>> Thanks,
>>
>> Jose
>>
>> -----Mensaje original-----
>> De: [e-mail] [mailto:[e-mail]] En nombre de
>> Dan
>> Wing
>> Enviado el: viernes, 27 de julio de 2012 0:53
>> Para: 'Jose Saldana'
>> CC: [e-mail]
>> Asunto: Re: [tcmtf] [tsvwg] M2M, WSNs and draft-saldana-tsvwg-tcmtf
>>
>>> -----Original Message-----
>>> From: Jose Saldana [mailto:[e-mail]]
>>> Sent: Thursday, July 26, 2012 3:56 AM
>>> To: 'Dan Wing'
>>> Cc: [e-mail]
>>> Subject: RE: [tcmtf] [tsvwg] M2M, WSNs and draft-saldana-tsvwg-tcmtf
>>>
>>> Dan,
>>>
>>> I don't know if I have really understood your question. Of course, I
>>> assume that a delay of some seconds is not important for certain
>>> sensor networks.
>>> But they don't need to signal that they are delay tolerant. I mean,
>>> they are assumed to be from the moment they are installed.
>> The device, doing the multiplexing, needs to know if the packets belong
>> to a
>> delay-tolerant (sensor) endpoint or to a non-delay tolerant endpoint.
>>
>> If the device doing the multiplexing is also the gateway communicating
>> with
>> the delay-tolerant (sensor) network, it certainly knows those
>> devices are delay-tolerant.   I believe that is your point.
>>
>> My point is that if the device doing the multiplexing is not directly
>> connected to that network of delay-tolerant devices, it needs to
>> somehow
>> differentiate between the delay-tolerant devices and the non- delay-
>> tolerant
>> -- by being signaled or by being configured.
>>
>> In the future, I expect a mix of sensors on various networks that are
>> not
>> traditional dedicated sensor networks (e.g., WiFi, 3G), so relying on a
>> gateway to that dedicated network may not always be the way to know a
>> device
>> is delay tolerant.
>>
>> -d
>>
>>> Well, I think I should gather some information about typical examples
>>> of sensor networks (number of sensors, frequency of samples, packet
>>> size of the samples, etc), in order to calculate in which cases the
>>> bandwidth saving obtained with TCMTF can be interesting, and which
>>> "multiplexing period"
>>> should be used. I think that this could be a typical case in which
>>> multiplexing can be done by packet size instead of using a period. I
>>> mean, when you have almost 1500 bytes (or the MTU), you send a
>> packet.
>>> I don't know if I have answered your question,
>>>
>>> Jose
>>>
>>> -----Mensaje original-----
>>> De: Dan Wing [mailto:[e-mail]]
>>> Enviado el: martes, 24 de julio de 2012 23:58
>>> Para: [e-mail]
>>> CC: [e-mail]; [e-mail]; [e-mail]
>>> Asunto: Re: [tcmtf] [tsvwg] M2M, WSNs and draft-saldana-tsvwg-tcmtf
>>>
>>> ...
>>>> Nevertheless, in order to get an idea of the savings, we should
>>> deploy
>>>> some calculations. Our research group has only done simulations for
>>>> VoIP and games. In this case, we should know:
>>>>
>>>>
>>>>
>>>> -          Number of sensors in a typical scenario.
>>>>
>>>> -          Packet size and used headers. If IPv6 is used, savings
>>>>            can be really big, since header compression is really
>>>>            good for it.
>>>>
>>>> -          Packets per second that a single sensor can generate.
>>>>
>>>>
>>>>
>>>> And of course, the gateway should be the place for multiplexing.
>>>> In this case, the flows only travel from sensors to data center.
>>>> Am I right?
>>> I wonder if there is an easy way for the endpoints to signal that
>> they
>>> are delay tolerant (to allow collecting ~5 seconds of little
>> packets).
>>> -d
>>
>> _______________________________________________
>> tcmtf mailing list
>> [e-mail]
>> https://www.ietf.org/mailman/listinfo/tcmtf
>
>