Re: [tcpm] [tsvwg] Agenda requests for TSVWG@IETF101

["Scharf, Michael (Nokia - DE/Stuttgart)" <[e-mail]>]

If somebody is interested in exploring this, I suggest to have a look at http://www.ikr.uni-stuttgart.de/Content/Publications/Archive/Sf_Diss_40112.pdf (Section 4.5, and a lot of the rest of the document). This document discusses how to process Quick-Start in routers. I never found the time to publish some of the content elsewhere.

In a nutshell, there can be different ways how a router could process Quick-Start requests with tradeoffs between performance, fairness, and risk of congestion.

However, the fundamental problem is more how applications can indeed make a reasonable request for bandwidth. Quick-Start works relatively well if it is enabled for these applications that indeed leverage it (e.g., mid-sized request-response traffic patterns), and if the requested rates are reasonable.

The Quick-Start mechanism basically fails completely e.g. if Quick-Start requests are also included in short TCP data transfers that complete in one or two RTTs. In that case, if the routers keep track of the request they approve, a lot of capacity will be kept for flows that will actually never use the rate they have requested, at the cost of other TCP connections that may be able to use it. There is also a fairness issue then. This issue can be solved by a more optimistic approval scheme that assumes that not all requests will indeed use the full capacity. But then one risks congestion as one over-commits capacity. Due to all these issues, it is non-trivial to find realistic scenarios in which Quick-Start indeed outperforms end-to-end congestion control e.g. with CUBIC and IW 10...

More below...

> Maybe I'm  saying what is obvious, but to be sure:
> 
> The way I see QS is:
> - The router tracks the unutilised link capacity at the current time, and uses
> this as some estimate of what may be available to new flows in the near
> future.

Even knowing the unutilized link capacity is a HUGE problem of its own. Even for Ethernet router ports, if there are VLAN-tagged ports, there is no such thing like a single unutilized link capacity... The Ethernet port is shared between different routing interfaces. And let's not look at wireless...

> - The router responds to QS requests and can accept them if there is
> "unutilised capacity" remaining.
> - When accepted router updates it's count of the total unutilised capacity
> reducing by the accepted QS requests. (i.e., Subsequent QS requests in the
> same interval will currently have a smaller pool of unutilised capacity
> available).

Indeed, that is the overall principle, but within that concept there are a number of different designs possible.

> - After the end of the interval all "QS requests" expire and the router again
> makes a measure of the unutilised capacity at the current time.

That concept of an interval also has issues. In general, the mechanism will work if the interval is of the same order of magnitude like the RTT of most connections using Quick-Start. If the RTT distribution is very heterogeneous, it is non-trivial how to let QS requests expire. One might need to create per-QS request state in the router, which has huge scalability issues in the fast path. 

> None of this implies a "reservation" of link resource in my thinking. In QS, the
> whole link capacity is intended to be shared using congestion control, and if
> flows start to send more using normal CC within the interval where QS is
> operating, they would simply receive a larger share of the capacity. That is:
> QS request didn't "allocate" any link resource.

Indeed, but some potential algorithms in a QS router could come close to "reservation" schemes. And then one runs basically into the same issues like IntServ.

Michael

 
> Gorry
> 
> 
> On 13/03/2018, 06:58, Katsushi Kobayashi wrote:
> > Hi,
> >
> > I think TCP QS router on the path also reserves bandwidth resource, if
> > it accepts a request.
> >
> > RFC4782 says:
> >    If the router approves the Quick-Start Request, this approval SHOULD
> >    be taken into account in the router's decision to accept or reject
> >    subsequent Quick-Start Requests (e.g., using a variable that tracks
> >    the recent aggregate of accepted Quick-Start Requests).
> >
> >
> > ---
> > Katsushi Kobayashi
> >
> >> 2018/03/13 午後3:43、Yingzhen Qu <[e-mail]
> >> <mailto:[e-mail]>> のメール:
> >>
> >> Hi Michael,
> >> Sorry for the late response.
> >>
> >> Thanks for pointing out that we missed this important reference.
> >> You're right, Quick-Start and our proposal do have lots of
> >> similarities, for example both of them require that end-points and
> >> routers to work together. But they are also different in details. For
> >> example, in our proposal in-band signaling proposal bandwidth is
> >> reserved on routers along the path.
> >>
> >> In next version of this draft, We'll add discussions about RFC 4728
> >> and 6077.
> >>
> >> BTW, can I request a slot to present this draft in TCPM if time allows?
> >>
> >> Thanks,
> >> Yingzhen
> >>
> >> On Mon, Mar 5, 2018 at 12:09 AM, Scharf, Michael (Nokia -
> >> DE/Stuttgart) <[e-mail]
> >> <mailto:[e-mail]>> wrote:
> >>
> >>     I believe that this proposal is similar to QuickStart TCP (RFC
> >>     4782), which is not cited in draft-han-tsvwg-cc, and the
> >>     reference is also missing in
> >>     draft-han-6man-in-band-signaling-for-transport-qos.
> >>
> >>     RFC 6077 explains some of the issues that an in-band signaling
> >>     mechanism like Quick-Start has to solve. As far as I can tell,
> >>     the fundamental challenge is neither the protocol specification
> >>     nor a prototype implementation. For instance, it has been proven
> >>     that QuickStart TCP can be implemented e.g. in network processors
> >>     (see
> >>     http://www.ikr.uni-
> stuttgart.de/Content/Publications/Archive/Sf_Diss_40112.pdf
> >>     <http://www.ikr.uni-
> stuttgart.de/Content/Publications/Archive/Sf_Diss_40112.pdf>).
> >>
> >>     So, when updating the documents, I suggest to add a discussion of
> >>     how the open research issues explained in RFC 6077 are addressed.
> >>
> >>     Michael
> >>
> >>     *From:* Yingzhen Qu [mailto:[e-mail]
> >>     <mailto:[e-mail]>]
> >>     *Sent:* Sunday, March 04, 2018 9:59 PM
> >>     *To:* [e-mail] <mailto:[e-mail]>
> >>     *Cc:* Scharf, Michael (Nokia - DE/Stuttgart)
> >>     <[e-mail] <mailto:[e-mail]>>;
> >>     Thomas Nadeau <[e-mail]
> >>     <mailto:[e-mail]>>; [e-mail]
> >>     <mailto:[e-mail]>; Yingzhen Qu <[e-mail]
> >>     <mailto:[e-mail]>>; [e-mail]
> >>     <mailto:[e-mail]>; [e-mail]
> >>     <mailto:[e-mail]>
> >>     *Subject:* Re: [tsvwg] Agenda requests for TSVWG@IETF101
> >>
> >>     Hi Gorry and Michael,
> >>
> >>     Thanks for the suggestion, I'll request a presentation in ICCRG.
> >>     Meanwhile, I think since the in-band signaling draft was
> >>     presented in TSVWG, if time allows it still makes sense to
> >>     present this draft in TSVWG.
> >>
> >>     The in-band signaling draft covers lots of aspects, and the
> >>     required changes include network layer and transport layer. We're
> >>     working on updating the draft, and may break it into pieces to
> >>     fit different WGs.
> >>
> >>     Your comments and help are very much appreciated.
> >>
> >>     Thanks,
> >>
> >>     Yingzhen
> >>
> >>     On Sun, Mar 4, 2018 at 4:49 AM, Gorry Fairhurst
> >>     <[e-mail] <mailto:[e-mail]>> wrote:
> >>
> >>         I am unsure yet what the correct group of people world be to
> >>         explore a "Bandwidth Guaranteed Network". The presentation
> >>         last IETF looked like the work could imply a need for changes
> >>         proposed to the network layer (using OAM exchnages) to set
> >>         the sending rate and make those bandwidth reservations.  In
> >>         the end, it could result in a protocol quite different to
> >>         TCP, I think this sort of change may possibly have a home in
> >>         TSVWG  - but first I'd agree with Michaeland would encourage
> >>         a presentation of the problem statement in ICCRG to explore
> >>         the issues.
> >>
> >>         Gorry
> >>
> >>         On 04/03/2018, 10:34, Scharf, Michael (Nokia - DE/Stuttgart)
> >>         wrote:
> >>
> >>
> >>             Hi all,
> >>
> >>             >From the abstract: “…This draft proposes a new TCP
> >>             congestion control algorithm used in bandwidth guaranteed
> >>             networks.  It is an extension to the current TCP standards.”
> >>
> >>             In the IETF, I believe the expertise for this specific
> >>             document would be in TCPM, which in CC. If the authors
> >>             are interested in feedback on the proposed mechanism, I
> >>             would recommend to ask TCPM.
> >>
> >>             Alternatively, corresponding research could perhaps be
> >>             performed in the ICCRG. ICCRG has published RFC 6077 to
> >>             document some of the open research issues in this space.
> >>
> >>             Michael
> >>
> >>             *From:*tsvwg [mailto:[e-mail]
> >>             <mailto:[e-mail]>] *On Behalf Of *Yingzhen Qu
> >>             *Sent:* Sunday, March 04, 2018 6:55 AM
> >>             *To:* [e-mail] <mailto:[e-mail]>;
> >>             [e-mail] <mailto:[e-mail]>
> >>             *Cc:* Thomas Nadeau <[e-mail]
> >>             <mailto:[e-mail]>>
> >>             *Subject:* [tsvwg] Agenda requests for TSVWG@IETF101
> >>
> >>             Dear Chairs,
> >>
> >>             A new draft (The draft was suggested by TSVWG @IETF100)
> >>             was just submitted, and we’d like to request a time slot
> >>             to present it @IETF101.
> >>
> >>             Title:A New Congestion Control in Bandwidth Guaranteed
> >>             Network
> >>
> >>             Presenter: Yingzhen Qu (Huawei)
> >>
> >>             Time required (including Q/A): 10 mins
> >>
> >>             Draft:
> >>             https://datatracker.ietf.org/doc/draft-han-tsvwg-cc/
> >>             <https://datatracker.ietf.org/doc/draft-han-tsvwg-cc/>
> >>
> >>             If there is any question, please kindly let us know.
> >>
> >>             Thanks,
> >>
> >>             Yingzhen
> >>
> >>
> >