Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

Re: [RPRWG] Complete RPR/Java model for all three proposals




Siamack, Bob, Adisak and all,

I'd like to add a few comments to the discussion on performance and
simulations.

From Bob's note:
> The first class, are the patterns that push the
> boundaries of important classes of expected traffic.  I expect that the
> output of the performance committee focused on this case.

In the performance committee, we discussed and concluded that it is
important to see the effects of non-uniform traffic distributions, as well
as the effects of bursty traffic on the metrics defined by the committee
such as fairness, bandwidth utilization efficiency. These can demonstrate
important performance characteristics of various proposals that we need to
pay attention to while establishing an RPR standard. This is because
burstiness can throw a closed-loop control system, which RPR essentially is,
into instability and/or cause oscillation. We certainly don't want these
type of problems to show up in the field after the standard is established.
Has anyone looked into these effects?

From Siamack's note:
> I would like to add that the likelihood of occurrence of such traffic
scenarios
> is also of utmost importance.  Usually such pathological traffic patterns
> are rare events, i.e. far towards the tail of the distribution.

That would be nice. However, although this statement makes a good point, we
can't easily predict the traffic patterns that RPR implementations will
encounter. Assuming that current Internet traffic patterns is all RPR will
need to deal with would be unrealistic. We also can't tell at this point all
the applications that RPR will be used for. I believe that we should
investigate the performance of various proposals under lenient and tough
conditions. This way we can make an informed decision by selecting the one
that gives the best overall results.

Best Regards,

Khaled Amer
President, AmerNet Inc.
Architecture Analysis and Performance Modeling Specialists

Address:     13711 Solitaire Way, Irvine, CA 92620
Phone:        (949)552-1114                      Fax:     (949)552-1116
e-mail:         amer@xxxxxxxxxxx
Web:           www.performancemodeling.com



----- Original Message -----
From: "RDLove" <rdlove@xxxxxxxxx>
To: "Adisak Mekkittikul" <adisak@xxxxxxxxxxxxxx>; "'Siamack Ayandeh'"
<sayandeh@xxxxxxxxxx>
Cc: "Khaled Amer" <amer@xxxxxxxxxxx>; "Stein Gjessing" <steing@xxxxxxxxx>;
<stds-802-17@xxxxxxxx>
Sent: Thursday, December 27, 2001 8:54 AM
Subject: Re: [RPRWG] Complete RPR/Java model for all three proposals


>
> Siamack and Adisak, you both make good points.  I believe I have a
slightly
> different point of view indicating a further area where we all need to be
> careful.  When defining how RPR works, it would be far better to have an
> algorithm that does not have holes in it that need to be explained away.
As
> an example look at the way IEEE 802.3 has consistently treated performance
> issues, such as distance limitations and maximum frame size.  The criteria
> they have always used is that their technology must work for the
absolutely
> worst case.  Any engineer can argue from a technical point of view that
this
> approach is too conservative.  However, from a marketing point of view
each
> weakness that needs to be explained with probabilities is a significant
> source of F.U.D. (Fear Uncertainty and Doubt) that can lead to decreased
> acceptance of our technology.  I strongly believe that we should attempt
to
> find a solution that has NO fundamental weaknesses, even if the likelihood
> of exercising that weakness is small.  Only if we can't achieve the above
> objective should we consider solutions that require explanations as to why
> the technology's fundamental weaknesses are really OK.
>
> If we expect RPR to be a huge success, then we can expect all present
> implementations to represent only a tiny fraction of RPR's market
> penetration.  In addition, the range of applications addressed today, is
> also likely to be only a small fraction of the applications that will
> eventually be handled.  Therefore, failure to see field problems today is
> not relevant to proving that similar algorithms are sufficiently robust.
>
> Best regards,
>
> Robert D. Love
> Chair, Resilient Packet Ring Alliance
> President, LAN Connect Consultants
> 7105 Leveret Circle     Raleigh, NC 27615
> Phone: 919 848-6773       Mobile: 919 810-7816
> email: rdlove@xxxxxxxx          Fax: 208 978-1187
> ----- Original Message -----
> From: "Adisak Mekkittikul" <adisak@xxxxxxxxxxxxxx>
> To: "'Siamack Ayandeh'" <sayandeh@xxxxxxxxxx>; "RDLove" <rdlove@xxxxxxxxx>
> Cc: "Khaled Amer" <amer@xxxxxxxxxxx>; "Stein Gjessing" <steing@xxxxxxxxx>;
> <stds-802-17@xxxxxxxx>
> Sent: Wednesday, December 26, 2001 2:10 PM
> Subject: RE: [RPRWG] Complete RPR/Java model for all three proposals
>
>
> >
> > Siamack,
> >
> >
> > I agree that it might be necessary to weight the problems we might find.
> > But at this phase of the standard development, I think we should be
> > more critical at finding the performance weaknesses of all proposed
> > algorithms.
> >
> > We need have all necessary data points to make a good decision. So far
> > the performance evaluation has been more of marketing talk. It's
> > mostly about how well this algorithm and that algorithm perform.
> > That is not good for our selection criteria.
> >
> > We need to do more of stress testing to determine the limitation of each
> > proposed algorithm. There are a few basic tests we can start with:
> >
> > 1. Control loop convergence test
> > 2. Fairness index
> > 3. Loss and delay
> > 4. Link Utilization
> > 5. Spatial reuse
> >
> >
> > Although coming up with good stressful scenarios to test the above
metrics
> > can be hard, I think it's achievable in a short time frame if we get
> people
> > to contribute the "worst" traffic pattern that they know of.
> >
> > What do you say folks? Any volunteer to compile the list?
> >
> >
> >
> > Best Regards,
> > Adisak
> >
> >
> >
> >
> >
> >
> >
> > -----Original Message-----
> > From: Siamack Ayandeh [mailto:sayandeh@xxxxxxxxxx]
> > Sent: Friday, December 21, 2001 8:18 AM
> > To: RDLove
> > Cc: Amer, Khaled; Stein Gjessing; stds-802-17@xxxxxxxx
> > Subject: Re: [RPRWG] Complete RPR/Java model for all three proposals
> >
> >
> > Bob,
> >
> > Thanks for sharing your insightful experience re pathological traffic
> > scenarios.
> > In response to your suggestion:  "One of the criteria we should have in
> > choosing
> > an algorithm is that the
> > worst case pattern simulations exceed some minimum threshold."
> >
> > I would like to add that the likelihood of occurrence of such traffic
> > scenarios
> > is also of utmost importance.  Usually such pathological traffic
patterns
> > are
> > rare events, i.e. far towards the tail of the distribution.  Having this
> > likelihood would put the scenario in perspective and also determine the
> > required
> > urgency of a resolution.  I think it is important to attach this
"weight"
> to
> > the
> > criteria when it comes to selection of alternative algorithms.
> >
> > Regards, Siamack
> >
> > RDLove wrote:
> >
> > > To:  802.17 Reflector, and Stein Gjessing
> > >
> > > Stein, Khaled, and all simulation experts, this email focuses on how
we
> > > choose
> > > traffic simulation patterns to run.  Drawing on my experience in 802.5
> > Token
> > > Ring Working Group, I believe there are two classes of simulation
> patterns
> > > that are critical.  The first class, are the patterns that push the
> > > boundaries of important classes of expected traffic.  I expect that
the
> > > output of the performance committee focused on this case.  The second
> > class
> > > are
> > > those cases that could "break"  the standard, i.e. reduce its
> performance
> > > below acceptable limits.  I have not seen much focus on this second
> class
> > of
> > > patterns, yet that is the class of patterns that could jeopardize the
> > > success of RPR.
> > >
> > > In the 802.5 Token Ring Working Group we discovered a "killer frame"
> that
> > > caused a buffer to overflow.  For a couple of meetings, addressing and
> > > solving the "killer frame" problem consumed the working group.  We
were
> > also
> > > very concerned about the press talking about token ring having a
"killer
> > > frame" problem, and what that would do to token ring sales.  The
problem
> > was
> > > finally solved by significantly increasing the required buffer size,
so
> > that
> > > no "killer frame" existed.
> > >
> > > Early coaxial Ethernet, which used "vampire connectors" had its
> equivalent
> > > of the "killer frame".  It turned out that if you inserted vampire
> > connector
> > > taps into the coaxial cable at a spacing that was a simple harmonic of
> the
> > > fundamental frequency, standing waves were created that caused
reception
> > > errors.  That problem was fixed by marking the coaxial cable
> periodically
> > > where you were allowed to tap into it.  The markings were spaced a
> > distance
> > > that was not a simple harmonic of the fundamental frequency, so that
> large
> > > standing waves would not be created.
> > >
> > > RPR must allocate bandwidth based on an algorithm that by its nature
> > cannot
> > > react instantaneously to the traffic requirements.  The feedback, or
> > > feedforward nature of any algorithm will cause a response that is
> > analogous
> > > to a circuit designed with a pole.  The circuit has a resonant
> frequency,
> > > "f" and a "Q" associated with it.  Depending on the parameters chosen,
f
> > and
> > > Q could be changed.  If the Q is high, then the circuit response to an
> > > excitation near frequency f would be dramatically different from the
> > > response to other excitations.
> > >
> > > By undersanding the physics of the feedback or feedforward algorithm
we
> > are
> > > proposing, we should be able to create the traffic patterns that
excite
> > our
> > > models at their point of resonance.  The requirement is that the model
> > does
> > > not break under that excitation.  In other words, the performance does
> not
> > > drop below acceptable parameters under "worst case" traffic.
> > >
> > > If a problem is found, there are two ways of addressing it. One is to
> > change
> > > the point of resonance.  This is a poor change, since it only means
that
> > the
> > > present worst case pattern doesn't cause the problem, but a new
pattern
> > will
> > > uncover the problem again, as bad as it always was.
> > >
> > > The other change is to change the Q of the model such that the
> degradation
> > > in performance of the model at its point of resonance is still within
> > > acceptable bounds.
> > >
> > > Clearly, each model will have its own "worst case patterns".
Hopefully,
> > > those
> > > groups forwarding proposals will understand their algorithms well
enough
> > > to be able to supply the worst case simulation patterns for that
model.
> > >
> > > One of the criteria we should have in choosing an algorithm is that
the
> > > worst case pattern simulations exceed some minimum threshold.
> > >
> > > I believe that for RPR to be successful we must require acceptable
> > > performance levels for worst case patterns. I assume that whatever
> > > algorithms are incorporated into our draft in January will be
thoroughly
> > > simulated before the March meeting to be sure there are no "killer
> > patterns"
> > > for RPR.   I would be pleased to hear feedback from the simulation
> experts
> > > on the need to include this criteria as one of the bases for adopting
a
> > > solution.
> > >
> > > Best regards,
> > >
> > > Robert D. Love
> > > Chair, Resilient Packet Ring Alliance
> > > President, LAN Connect Consultants
> > > 7105 Leveret Circle     Raleigh, NC 27615
> > > Phone: 919 848-6773       Mobile: 919 810-7816
> > > email: rdlove@xxxxxxxx          Fax: 208 978-1187
> > > ----- Original Message -----
> > > From: "Stein Gjessing" <steing@xxxxxxxxx>
> > > To: <stds-802-17@xxxxxxxx>
> > > Sent: Tuesday, November 27, 2001 8:54 AM
> > > Subject: [RPRWG] Complete RPR/Java model for all three proposals
> > >
> > > >
> > > > All,
> > > >
> > > > I will try to implement all three proposals:
> > > >
> > > > Alladin,
> > > > DVJ,
> > > > Gandalf
> > > >
> > > > in my RPR/Java-simulator and get some simple traffic
> > > > scenarios running before the January meeting.
> > > >
> > > > I have the early November drafts of all these proposals as they are
> > > > posted on the web site. However, I need working drafts of more
> > > > details of the fairness/flow control algorithm parts of these
> > > > proposals. I talked to some of you during the Austin meeting
> > > > and you said you could send me material that would help me in
> > > > my Java implementation effort. So please do that now.
> > > >
> > > > Please make specifications as precise and algorithmic (step by
> > > > step description on what to do in different cases) as possible.
> > > >
> > > > All Java code (that is all the RPR/Java models including the
simulator
> > > > code itself) will be available for everyone.  If the material you
send
> > > > me contains confidential information (not related to the specific
> > > > fairness/flow control algorithm), I can treat that material
> > > > confidentially.
> > > >
> > > > I also suggest you send me traffic scenario proposals
> > > > (e.g. two for each of the Alladin, DVJ and Gandalf proposals)
> > > > that you think I should run.
> > > >
> > > > I hope to be able to do this during December, but depending
> > > > on how difficult this will be, as well as what else I have
> > > > to do in my job, I can of course not promise anything.
> > > >
> > > > Stein
> > > >
> > > >
> > > >
>
>