RE: [RPRWG] RPR Perf: Single vs. Dual ring?]
Well, this is not about DPT/SRP. We should focus on fundamental issue
of defining RPR MAC requirements. I think, besides fainess, these
following requirements should be given a serious consideration before we
begin to evaluate and compare each proposal.
1) the responsiveness: ring delay, shortest path
2) the robustness: fault tolerance, node failure, packet loss, etc
3) overhead: 1%, etc
4) scalibility: number of rings,<2, =2, > 2; etc.
Adisak
-----Original Message-----
From: Donghui Xie [mailto:dxie@xxxxxxxxx]
Sent: Monday, April 30, 2001 4:04 PM
To: Adisak Mekkittikul
Cc: stds-802-17@xxxxxxxx
Subject: RE: [RPRWG] RPR Perf: Single vs. Dual ring?]
Adisak,
In a 10G RPR ring, the flow control traffic overhead should be much less
than 1% of the bandwidth. Any practical scenario can justify such a "highly
desirable" shortest path that it wouldn't allow the sub 1% control
flow? What are the superiorities for a flow control approach to have to
route the sub 1% flow control packet to go the longest path around the
whole ring? One thing that I can see is that such a "superiority" approach
would maximize the sub 1% flow control interference with data flows,
instead of minimize it.
If you haven't learned DPT/SRP IPS's full protection and fault tolerance, I
suggest you take some time and read RFC2892.
Donghui
At 02:22 PM 4/30/2001 -0700, Adisak Mekkittikul wrote:
>Donghui,
>
>I agree that using the opposite ring in one case can give a shortest path (
>although still ring size dependent-- O(n)). However, if the shortest path
>is highly desirable, using both rings (clockwise and counter-clockwise) can
>be
>a more superior approach since it definitely guarantees the shortest path
in
>
>all cases.
>
>BTW, can you comment on fault the tolerance of a hop-by-hop scheme in the
>event that
>one or more intermediate node fail?
>
>Adisak
>
>
>
>-----Original Message-----
>From: Donghui Xie [mailto:dxie@xxxxxxxxx]
>Sent: Monday, April 30, 2001 10:37 AM
>To: Stein Gjessing
>Cc: stds-802-17@xxxxxxxx
>Subject: Re: [RPRWG] RPR Perf: Single vs. Dual ring?]
>
>
>
>Hi Stein,
>
>Further with your answers, in dual rings, backward hop-by-hop control
>packet gives the fastest response time that is possible, since the control
>packet is allowed to go the shortest path to reach the node. As such, the
>flow control is less dependent on ring global size, and helps to minimize
>the interference with data flows. On the other hand, a broadcast control
>packet in single ring has to go the longest path, which not only makes the
>flow control be more sensitive to the ring size and distance, but also
>subjects the data flows and control flows over the whole ring to mutual
>interference.
>
>Donghui
>
>At 03:54 PM 4/28/2001 +0200, Stein Gjessing wrote:
>
> >Samian,
> >
> > > So, to clarify my understanding, we are using
> > > dual rings for flow control packets that propagate hop-by-hop
upstream.
> >
> >Yes.
> >
> >I believe there already is a (close to) consensus on dual rings.
> >Remember that dual rings are also needed for protection.
> >
> >The flow control packets are, as you say, in a sense broadcasted.
> >But on a ring a broadcast needs to go hop by hop.
> >And upstream is the sensible way to send the packets, because
> >the node that bugs you is most probably closest upstream.
> >
> > > But, wouldn't it make the simulations easier if we could default to a
>more
> > > generic flow control scheme,
> >
> >I believe we have to simulate and understand the flow control schemes
> >down to the very details. Small changes in flow control can make big
> >changes in performance. Flow control is feedback, and if we don't get
> >it right we can get oscillations etc., that are typical for a badly
> >controlled feedback system.
> >
> >Stein
> >
> >
> >------Original Message-----
> >Return-path: <owner-stds-802-17@xxxxxxxx>
> >Envelope-to: steing@xxxxxxxxxx
> >Delivery-date: Sat, 28 Apr 2001 02:30:16 +0200
> >From: Samian Kaur <skaur@xxxxxxxxxxxxxx>
> >To: "'stds-802-17@xxxxxxxx'" <stds-802-17@xxxxxxxx>
> >Subject: Re: [RPRWG] RPR Perf: Single vs. Dual ring?
> >Date: Fri, 27 Apr 2001 16:43:50 -0700
> >MIME-Version: 1.0
> >X-Mailer: Internet Mail Service (5.5.2653.19)
> >Content-Type: text/plain;
> > charset="iso-8859-1"
> >Sender: owner-stds-802-17@xxxxxxxx
> >Precedence: bulk
> >X-Resent-To: Multiple Recipients <stds-802-17@xxxxxxxxxxxxxxxxxx>
> >X-Listname: stds-802-17
> >X-Info: [Un]Subscribe requests to majordomo@xxxxxxxxxxxxxxxxxx
> >X-Moderator-Address: stds-802-17-approval@xxxxxxxxxxxxxxxxxx
> >
> >
> >
> >Hi Stein,
> >
> >Thank you for the explanation. So, to clarify my understanding, we are
>using
> >dual rings for flow control packets that propagate hop-by-hop upstream.
> >
> >But, wouldn't it make the simulations easier if we could default to a
more
> >generic flow control scheme, like maybe broadcasting the control packets.
> >This will be a more robust mechanism than hop-by-hop and also eliminate
the
> >necessity of the second ring. The performance characteristics of this
>scheme
> >will also give us an insight into the performance tradeoff of this
approach
> >for comparison.
> >
> >Do you think a scenario on these lines will be interesting?
> >
> >Thanks,
> >Samian
> >
> >- -----Original Message-----
> >From: Stein Gjessing [mailto:steing@xxxxxxxxxx]
> >Sent: Tuesday, April 24, 2001 11:41 PM
> >To: stds-802-17@xxxxxxxx
> >Subject: [RPRWG] RPR Perf: Single vs. Dual ring?
> >
> >
> >
> >Samian,
> >
> >I agree that a single ring will demonstrate the performance
characteristics
> >of interest if we disregard flow control.
> >Analyzing performance with flow control (by sending flow control
> >packets upstream), the dual rings are of course necessary.
> >Then the interference between flow control packets and data packets
> >also become an interesting issue.
> >
> >(Also remember that packets are going at most half way around on one
ring)
> >
> >Stein Gjessing
> >University of Oslo
> >
> >
> > >Hi Khaled,
> > >
> > >I know that the performance adhoc committee decided that the Phase I
> > >simulations should be done using dual rings. I am beginning to question
>if
> > >that is necessary. I think a single ring will demonstrate the
performance
> > >characteristics of interest and save us a lot of run time in running
the
> > >simulations and is much easier to analyze.
> > >
> > >Am I missing something?
> > >
> > >Samian Kaur
> > >Lantern Communications