|
Necdet, thank you for supplying values that would
be of general interest.
I have one further question for you. You have
indicated in your note "This would give us
how much of a buffering needed ..." I thought that Jon indicated that
additional buffering may not help, it will just change the latency / traffic
load on the ring - conditions for which the problem occurs. If Jon's
simulations with your recommended values confirm this hypothesis, would you say
that we have a problem that needs correction?
I'm trying to pin down what simulations with what
results would indicate a problem because Jon's hypothesis is that we have a
problem. Until we can get agreement on what simulations would suggest we
have a condition that needs fixing, we are likely to just end up arguing about
data, rather than about what problems we may have and what it will take to fix
them.
Thank you again Necdet for your assistance in
carefully defining the conditions to be simulated, and the results that would
indicate changes are required.
Best regards,
Robert D. Love 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 -----
Sent: Tuesday, April 22, 2003 8:08
PM
Subject: Re: [RPRWG] Class A and B
Guarantees
Bob,
I would like to see simulations for the scenario that Jon described with
the following conditions:
stqLowThreshold is fixed to a value (say 100kB)
stqHighThreshold is fixed to a value (say 200kB)
stqFullThreshold is
set to infinity (or to a very large value, say 100MB)
head node is adding
classA1 traffic at 10% of line rate.
It would be nice to see the maximum stq buffer occupancy in the head node
with respect to number of nodes on the ring. This would give us how much of a
buffering needed in order not to hit the stqFullThreshold.
This result can also be used as a check mechanism for the formulas that
Annex G editor(s) provided.
Thanks.
Necdet
"Robert D. Love" wrote:
Necdet, with
regards to my question: RDL: Are there any conditions that would lead you to conclude there
is a potential problem with our present algorithm?
...and your answer
NU: No.
But, I want to make sure that things are clear to all of us.
Let me first apologize for not making my
question clear enough, and now let me try again.What I would like to know, and what I believe would
be most helpful to Jon as he runs his simulations is the
following: What initial
conditions must Jon use in his simulations so that you will agree that a
valid simulation with these conditions produces a meaningful result.
And then, what result with those initial conditions, would have you in
agreement that we have a problem (assuming the simulation was done
correctly). Necdet, I want
to avoid having a running argument where whatever is simulated is
challenged. If we are to make good progress on this issue, we need
your input as to what initial conditions need to be set, so that you will
not be challenging those conditions. If Jon or others believe that
other initial conditions should be used, then let's have a dialog about
which conditions need simulation, rather than focusing arguments on
challenging results because we disagree on the initial
conditions. I am hoping
that you, Jon, and other simulation experts can work as a team in
establishing those runs we need to evaluate, and in agreeing, in advance,
what types of results would indicate the algorithms we are using have a
problem. - Of course, if the algorithms have no problem, then
those agreed to simulations should not produce any alarming
results. Thank you
Necdet. Best regards, Robert D. Love
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 -----
Sent: Tuesday, April 22, 2003 4:39
PM
Subject: Re: [RPRWG] Class A and B
Guarantees Bob,
"Robert D. Love" wrote:
Necdet,
please fill us in by adding a bit more verbiage. i.e. Why are you
asking these particular questions?
NU: I am trying to understand the
scenarious that he mentioned so that we can speak the same
language.
What are the implications of a yes
response, of a no response?
NU: Yes, for example not having a
shaperD would not provide any guarantees for classA0
traffic.
Are there particular conditions that
Jon should be simulating?
NU: If he thinks that there are
scenarious that we have not looked at, we need to look at them. However,
so far I have not seen anything in his e-mail that we have not looked
at.
Are there any conditions that would
lead you to conclude there is a potential problem with our present
algorithm?
NU: No. But, I want to make sure that
things are clear to all of us.
Thank you Necdet.
Best regards,
Robert D. Love
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 -----
Sent: Tuesday, April 22, 2003
2:24 PM
Subject: Re: [RPRWG] Class A and
B Guarantees Jon,
Please see my comments in line.
Thanks.
Necdet
Jon Schuringa wrote:
Dear all,
I posted a comment (#33) at the
Dallas meeting about bandwidth
guarantees: In my opinion, bandwidth
agreements cannot always
be guaranteed. The comment was rejected because it was addressed to the
wrongclause. Although at the wrong address, I got the answer that
thestatement in my comment is incorrect, but without any
explanation. Since then
I had discussions with several people, and checked my
simulations with another
simulation tool (ns2). As before, I strongly
believe this to be a serious technical
concern, and therefore post it
here to the mailing list.
The problem in
short:
STQ's can
reach the stqFullThreshold in scenarios where both class
C
and class A
traffic flows. As a result, the STQ gets precedence
over
all locally
sourced traffic, so that class A (and B) traffic has to
wait,
causing
bandwidth and jitter problems.
The STQ can get that full because fairness
messages cannot stop
packets that already have been transmitted by other
stations, but did
not yet arrive at the local station. This amount of packets
that is on the
transit path can be very large since it is the sum of all
packets in the
STQs
on the transit path. This is also the reason why larger
STQs
do not solve
the problem. So
basically what happens in the problem scenarios is
that:
1) the
local station (S) receives class C packets at 100% of the line
rate.
All these packets need to be
forwarded by station S
2) Station S transmits guaranteed class A (local)
traffic at some rate x,
so the local STQ grows (at rate
x).
3)
Station S advertises a fair rate unequal to FULL_RATE once the
STQ
exceeds the
stqLowThreshold
4) All other stations see the advertized rate and
limit their "add" traffic.
This however does not directly prevent
that station S gets less than
100% line rate, because
there is still transit traffic that needs to be
forwarded by
all stations. These stations empty their STQs.
5) If the class A rate x and the
number of STQs are "large enough", the
STQ in station S will
reach its stqFullThreshold and priority inversion
is the result. Note
that the potential problem scenarios are realistic hub-scenarios,
not
"pathological
cases".
NU: Did you run any simulations
showing the priority inversion happening while adding classA1 (when
stqFullThreshold - stqHighThreshold > RTT * rateA1?
A detailed description and
an example scenario can be found here:
http://grouper.ieee.org/groups/802/17/member/draftballots/d2_1/refs/js_issues_1.pdf
This document contains other
issues as well.
Opinions?
NU: Did you implement shaperD and
reserved classA0 bandwidth all around the ring?
Best
regards,Jon -----------Jon SchuringaInstitute of Communication
Networks
Vienna
University of Technology
Favoritenstraße 9/388
A-1040 Vienna
+43/1/58801-38814
www.ikn.tuwien.ac.at
|