Internet DRAFT - draft-dudley-rtecn-simulation
draft-dudley-rtecn-simulation
TSVWG S. Dudley
Internet Draft Nortel
Expires: January 2, 2006 July 2005
Simulation of RT-ECN based Admission Control and Preemption
draft-dudley-rtecn-simulation-00.txt
Status of this Memo
By submitting this Internet-Draft, each author represents that any
applicable patent or other IPR claims of which he or she is aware
have been or will be disclosed, and any of which he or she becomes
aware will be disclosed, in accordance with Section 6 of BCP 79.
Internet-Drafts are working documents of the Internet Engineering
Task Force (IETF), its areas, and its working groups. Note that
other groups may also distribute working documents as Internet-
Drafts.
Internet-Drafts are draft documents valid for a maximum of six months
and may be updated, replaced, or obsoleted by other documents at any
time. It is inappropriate to use Internet-Drafts as reference
material or to cite them other than as "work in progress."
The list of current Internet-Drafts can be accessed at
http://www.ietf.org/ietf/1id-abstracts.txt.
The list of Internet-Draft Shadow Directories can be accessed at
http://www.ietf.org/shadow.html.
This Internet-Draft will expire on January 12, 2006.
Copyright Notice
Copyright (C) The Internet Society (2005).
Abstract
This document summarizes simulation results obtained from studies of
a measurement based admission control and preemption scheme using
Real-Time ECN semantics for SIP voice session setup.
Conventions used in this document
Some acronyms are used in this document for brevity and may refer to
general concepts.
Dudley Expires - January 2, 2006 [Page 1]
Simulation of RT-ECN based Admission Control and Preemption
SAC (Session Admission Control) may refer to either Admission
Control, Preemption or both as required by the text
TDM (Time Division Multiplexing) refers to traditional Voice
Switching networks
ECN (Explicit Congestion Notification) refers to either the
concept of explicit congestion notification or to the bit
field identified for use with Explicit Congestion Control
per RFC 3168.
RT (Real Time) refers to either Real-Time services or
applications.
Table of Contents
1. Introduction...................................................2
1.1 Motivation.................................................2
1.2 Approach...................................................3
2. Notes on Token Bucket Modifications............................5
3. Simulation Setup...............................................7
4. Qualitative Performance........................................8
4.1 Admission Control- Qualitative Analysis....................8
4.2 Preemption - Qualitative Analysis..........................9
5. Quantitative Performance......................................11
5.1 Admission Control - Quantitative Analysis.................12
5.2 Preemption - Quantitative Analysis........................14
5.3 Impact of Multiple Application Managers...................16
6. Conclusions...................................................18
Security Considerations..........................................19
References.......................................................19
Intellectual Property Statement........Error! Bookmark not defined.
Disclaimer of Validity.................Error! Bookmark not defined.
Copyright Statement..............................................20
Acknowledgments..................................................20
Author's Addresses...............................................20
1. Introduction
This report summarizes simulations carried out on measurement based
admission control and preemption based on Real-Time ECN semantics.
1.1 Motivation
The investigation of admission control and preemption mechanisms is
part of an effort to look at ways of managing session bandwidth on IP
networks. The need to manage bandwidth may come either from a need
to protect the network from traffic surges, re-route events, etc., or
it may come from a need to prioritize session traffic on links of
limited capacity. Real-Time ECN semantics have the advantage of
taking action at the endpoints, rather than inside the network and so
can avoid the need to track flows within the network. The general
Dudley Expires - January 2006 [Page 2]
Simulation of RT-ECN based Admission Control and Preemption
approach also provides a wide latitude in choice of semantics that
can permit independent action by endpoints or application managers at
the edge of the network. Simulations were carried out to take a more
detailed look at the capabilities, and issues, that would arise from
this approach.
One motivation for looking at admission control and preemption is the
need to provide preferential treatment to some sessions on the
network. Examples of networks where multiple precedence levels exist
could include the U.S. Department of Defense DSN network which has 5
levels of precedence but could also include, under special
circumstances, the existing TDM voice network where some endpoints
are sometimes given special treatment. For example, the current TDM
voice network, under periods of network stress or during emergency
conditions, provides preferential treatment for fire, medical and
emergency services endpoints. Emergency services, such as the ones
listed above are deliberately wired to the switch in a way that
guarantees that they are the last to be dropped when load shedding is
needed, are the first to get service when a switch recovers from a
failure, and can get admitted to the network when all non emergency
callers are denied. The transition from a TDM voice network to an
all VoIP network would currently involve losing these capabilities so
mechanisms which could provide an equivalent capability have value.
1.2 Approach
The approach taken for Real-Time ECN is summarized below. A more
detailed description of this approach can be found in “Congestion
Notification Process for Real-Time Traffic, draft-babiarz-tsvwg-
rtecn-03”, Feb 2005i, and “RTP Payload Format for ECN Probing, draft-
alexander-rtp-payload-for-ecn-probing”, May 2005ii. The semantics of
the scheme are tailored for session based traffic such as is created
by SIP, and the simulations summarized here used SIP endpoints.
The basic notion of protecting the network is to ensure that the call
arrival rate never permanently swamps the call departure rate. While
admission control affects the arrival rate, preemption affects
(increases) the departure rate.
Network devices meter traffic and, based on thresholds, signal their
level of congestion to the endpoints by setting bits in the ECN field
of RTP packets emitted by endpoints. The endpoints react to that
signaling to either make independent admission control decisions, or
to notify an application manager that mediates the decision making
process for session preemption. Decisions are made in a way that
guarantees that , traffic of lower precedence is denied or preempted
before traffic of higher precedence.
Admission control is based on having endpoints send small RTP packets
(probe packets) through the network during call setup. Probe packets
Dudley Expires - January 2006 [Page 3]
Simulation of RT-ECN based Admission Control and Preemption
use the same IP address and port number as the media packets which
would follow so the probing process was guaranteed to have the same
congestion conditions as the media which would follow. The decision
to admit is made at the endpoint, based on rules coded into the
device.
The admission decision is based on session precedence and level of
congestion. Two levels of congestion are defined. Routine sessions
(i.e. the lowest precedence session and the most common kind of
session) call setup terminates with a network failure condition if
either Level 1 or Level 2 congestion is indicated on the link during
probing. Higher precedence sessions terminate call setup if Level 2
congestion is detected and persists for at least 500 milliseconds.
If congestion clears before the 500 millisecond timer expires, the
session is admitted. The 500 millisecond interval gives the network
some time to clear congestion via its preemption mechanisms. There
is a trade-off in how long to wait for the network to clear before
making the final admission decision. Since it is possible, even
though unlikely, that the congestion is caused entirely by high
precedence sessions, the session is denied at some point in time if
congestion does not clear. (i.e. The choice of a 500 millisecond
interval is somewhat arbitrary.)
Probing continues until the call is answered. For routine calls, if
a Level 1 indication is received during that time, call setup
terminates with a network failure condition. For high precedence
sessions, if a Level 2 indication is received during that time, it
follows the sequence of activities associated with preemption. (i.e.
the high precedence session has already been granted admission to the
network and should only be removed when other sessions of the same
precedence are also being preempted.)
The preemption process begins with all of the endpoints whose traffic
traverses an affected link detecting Level 2 congestion. If an
endpoint experiences congestion for a sufficiently long time, it
sends a NOTIFY message to an application manager that contains
contact information for both ends of the session. The length of time
that it waits before sending a NOTIFY message is randomly selected
from within a range dependent on the precedence of the session. Each
precedence level has a separate time range, separated by a minimum
length of time from other precedence ranges. The deliberate
separation of notification intervals permits all sessions of a lower
precedence to have their notifications reach the application manager
before notifications coming from higher precedence sessions, thus
ensuring that all session of lower precedence will become targets of
preemption before sessions of higher precedence.
In the implementation used for these simulations, the first NOTIFY
results in the application manager preempting the session (i.e it
sends BYE messages to both endpoints) and then setting a timer (500
Dudley Expires - January 2006 [Page 4]
Simulation of RT-ECN based Admission Control and Preemption
millisecond). No other preemptions are attempted until the timer
expires. During that interval, the application manager continues to
listen for congestion NOTIFY messages and builds a preemption list
from the messages.
The length of time that the application manager collects
notifications without between preemptions represents the expected
length of time it would take for the BYE message from the application
manager to reach endpoints, for them to stop sending media packets,
and for token buckets in the routers to decide whether the congestion
condition has changed. For a large network, with longer transmission
delays, it may need to be longer than 500 milliseconds, and for a
small network it may be shorter.
If a single preemption is sufficient to bring the congested link back
below the threshold, all endpoints will begin receiving a new
congestion indication. All endpoints that have sent a NOTIFY message
to the application manager send another NOTIFY message indicating
that congestion has cleared. The application manager removes them
from the preemption list.
Approach Notes:
In selecting which network elements need to meter and mark the ECN
field, it was noted that not all links on a given network can be
congested. Many, because of their relation to other links on the
network that always become congested first, can never become
congested. The full benefits of the admission control and preemption
scheme can be realized by implementing the metering and marking
process on a subset of those routers.
Management of the admission control and preemption decision at the
edges allows multiple behavior systems to be established and co-
exist. The initial simulation work described here only looks at
voice sessions established using SIP. Extension of this work to
include non-voice sessions (video conferencing) is underway. Lessons
learned in this type of investigation will likely have merit in
looking at other kinds of session based traffic.
2. Notes on Token Bucket Modifications
For use in the metering process in the router, a token bucket
algorithm is common and has many benefits for use in this scenario.
It provides the benefit of ensuring that the average threshold rate
has been exceeded for a sufficient time to reliably declare that the
threshold has been exceeded. When used with randomly arriving
packets it provides a very good random marking of a percentage of the
packets. However, for non-randomly arriving packets, such as those
coming form Real-Time sources, the marking behavior is also non-
Dudley Expires - January 2006 [Page 5]
Simulation of RT-ECN based Admission Control and Preemption
random and cannot be guaranteed to mark fairly across sessions of
different precedence.
Part of the problem arises from the non-statistical nature of the
inter-packet arrival times in a real-time flow like voice. Codecs
emit packets on a fixed interval to support the needs of a voice
application. When two codecs are emitting packets at the same fixed
interval, the relative arrival time of the two packets is quite
likely to be the same from frame to frame. This can cause a problem
with a system based on marking only when the bucket is empty. If the
spacing between packets is large enough in one frame to cause the
bucket to fill slightly, and thus not mark the packet as experiencing
congestion, it is quite likely to be so in the frame that follows as
well. This would mean that even though the average rate of all flows
exceeds the threshold, the particular flow that is lucky enough to
have a long interval between it and the previous packet will almost
always be marked as not experiencing congestion. The more desirable
pattern is to start marking when a congestion condition is first
detected and then not stop marking until we are certain that the
condition has cleared.
The mechanism chosen to modify the token bucket involved creating a
two state marking process. Starting from the un-congested condition,
no marking of packets is done. When the token bucket is emptied for
the first time, marking of packets begins. Marking of packets
continues until the token bucket fill level exceeds a defined
threshold. At that time, the marking of packets stops, and the token
bucket reverts back to its original state.
The size of the token bucket, or alternatively its burst size,
determines how quickly the token bucket responds to changes in
throughput. A small burst size responds quickly to changes, whereas
a large burst size requires a longer period of time to fill or empty.
In creating the two state device, it was noted during initial
simulations that if the burst size used for declaring the start of
marking and the burst size used for declaring the end of marking were
the same, that short burst sizes led to a risk of stopping the
marking process too early and large burst sizes led to the problem of
admitting more sessions before the endpoints start to receive
indication of congestion.
The solution to the issue was to make the burst size different for
detecting when to start marking than for detecting when to stop
marking. A small burst size for detecting the start of marking and a
much larger burst size for detecting when to stop marking produced a
very stable system, capable of reacting quickly to rapid onset events
but yet not incorrectly declaring drops in throughput. The mechanics
of how this was accomplished won’t be addressed in this report but
can be obtained by consulting with the author.
Dudley Expires - January 2006 [Page 6]
Simulation of RT-ECN based Admission Control and Preemption
3. Simulation Setup
The way that this simulation was undertaken was to take an existing
simulation program (OPNET Modeler) and to extend the behaviors of the
models available in it. A router model was extended to be able to
meter traffic on its links and set ECN bits in packets traversing its
links. A workstation model was extended to monitor ECN bits and to
simulate important elements of SIP message behaviors including:
INVITE, ACK, UPDATE, PRACK, CANCEL, BYE, and NOTIFY. The INVITE and
non-INVITE Client and Server Transaction models as described in RFC
3261 and elements of the Pre-conditions model as described in RFC
3312 were also implemented. The same workstation model was extended
in a different way to act as a Back to Back User Agent and as an
application manager to mediate preemption decisions.
The ECN metering process used by the router follows the
recommendations in “Congestion Notification Process for Real-Time
Traffic, draft-babiarz-tsvwg-rtecn-03”, Feb 2005.
The pre-conditions model as described in RFC 3312 was used as the
point of departure for the simulations. However, it was not followed
exactly as described in RFC 3312 since the semantics there pertain to
RSVP. At the point where RFC 3312 requires an indication of
permission to continue with call setup, the simulation triggers the
establishment of a flow of probe packets between the endpoints.
(More details about the probing process can be acquired from “RTP
Payload Format for ECN Probing, draft-alexander-rtp-payload-for-ecn-
probing”, May 2005.)
ECN bit marking on the probe packets is used to detect the level of
congestion on the network, which is used by the endpoint to make the
decision to proceed or to terminate call setup i.e. to carry out
admission control. After the call is established, the ECN markings
on RTP media packets are used to make preemption decisions.
The partition of functions was carried out as follows.
Who Meters Traffic Selected Routers on the network
Detection of congestion: metering of links using a modified token
bucket algorithm.
Who is signaled: SIP endpoints
How it is signaled: ECN bits are set with values 0,1 or 2
What is signaled Level 0 indicates no congestion, Level 1
indicates a first level of congestion and
Level 2 indicated a second level of
congestion. (No attempt was made to
distinguish between ECN capable and not
ECN capable flows or to mimic the exact
bit patterns of RT-ECN)
Dudley Expires - January 2006 [Page 7]
Simulation of RT-ECN based Admission Control and Preemption
Admit Action Endpoints make the decision based on the
precedence of their session as indicated
in the Approach section
Preempt Action Endpoints notify a Back to Back User Agent
(B2BUA), which fills the role of the
application manager described in the
Approach section, if congestion still
persists when an internal randomly
selected timer expires . The B2BUA
maintains a preemption list and sends BYE
messages to endpoints periodically (every
500 milliseconds).
4. Qualitative Performance
Before examining the performance of the system from a numerical
perspective, it is instructive to have a look at some of the
qualitative response of the system
4.1 Admission Control- Qualitative Analysis
The desired behavior of the admission control mechanisms is to deny
admission to routine or high precedence sessions whenever the
threshold for each is crossed on any link in the network, regardless
of where in the network this event occurs. This was the observed
behavior of the simulation. Getting the system to cross the second
threshold to test admission control of high precedence sessions,
however, proved to be difficult. Even with very high call arrival
rates, the second level of congestion could only ever be reached if
the volume of high precedence traffic, by itself, was sufficient to
generate enough traffic to exceed the second threshold.
This is not to say that traffic levels never exceed the first
threshold. There is some latency in the detection mechanism, both
coming from the token bucket filters, and from the latency of packet
traversal across the network that can allow more than one “extra”
call to be admitted above the Level 1 threshold. The closer together
that calls arrive, i.e. the higher the instantaneous call arrival
rate, the more sessions can be admitted before the congestion
indication is acted on at the edges. For routine calls, continuous
probing is very effective in reducing overshoot, even when call
arrival rates are very high.
After Level 1 ECN markings began to arrive at all endpoints, the
behavior of the system was to stop admitting routine calls. As will
be seen in the Quantitative Performance section of this document, the
system was able to tolerate very high call arrival rates. We have
not attempted to determine what call arrival rates are in line with a
mass calling event but the overall performance suggests that it is
Dudley Expires - January 2006 [Page 8]
Simulation of RT-ECN based Admission Control and Preemption
likely possible to engineer for protection against any level of mass
calling event.
The desired behavior of the admission control mechanism is to be
insensitive to where in the network the congestion occurs, and to how
many congestion events are present in the network. The results of
the simulation confirm this behavior. Marking of Level 1 and Level 2
congestion states by routers in the network was established so that
an incoming packet marked at Level 2 congestion from a previous link
would not be re-marked to indicate a Level 1 congestion. The
congestion indication received by the endpoint is always the worst
case congestion event anywhere in the network.
4.2 Preemption - Qualitative Analysis
Since the preemption control point (Level 2 Congestion indication) is
higher than the Level 1 congestion point, the network is already
denying all routine sessions by the time that preemption is required.
At that point, the arrival rate of high precedence calls has to
exceed the departure rate of all calls in order to reach Level 2
congestion. In the normal condition where high precedence calls are
a small percentage of total calls, that can be made highly unlikely
just by setting the Level 2 threshold far enough from the Level 1
threshold. In the core of a network where this condition can be
reasonably expected to be true, preemption would only be necessary to
handle the case of traffic re-route after a link failure. At the
edges of the network, it may be possible for a sudden surge to reach
a Level 2 congestion condition, and these are the cases that are
investigated in the Qualitative section that follows.
The desired behavior of the preemption process is to begin preempting
sessions when a specific threshold is reached in the network, to
remove only sessions that traverse the affected link, to always
remove sessions with a lower precedence before removing sessions of a
higher precedence and to stop removing sessions as soon as the
congestion level drops below the specified threshold. The observed
results conform to all of the requirements except for the effects of
latency in the start and stop conditions.
The latency comes in part from the token bucket which must either
empty or fill a certain amount before beginning to change ECN
markings, and in part from the transmission delay on the network.
The overall result is that the performance of the system at high call
arrival rates in slightly different than at low call arrival rates.
The differences, however, are small as will be seen later..
For the Level 2 threshold, since the preemption control point (Level
2 threshold) is also the admission control point for high precedence
traffic, the latency leads to a condition where high precedence calls
can be admitted for a short time after the Level 2 threshold is
Dudley Expires - January 2006 [Page 9]
Simulation of RT-ECN based Admission Control and Preemption
actually reached. Continuous probing does not have a significant
impact on reducing the number of high precedence sessions admitted
after the threshold is crossed because the behavior for high
precedence sessions that receive a Level 2 congestion indication
during ringing is based on the strategy of waiting to see if other
routine sessions could be preempted to allow them to remain
connected.
The mediation of the preemption process through the application
manager resulted in a small increase in signaling load. The increase
was kept small by to the randomization of the notification process.
This paced the delivery of NOTIFY messages. Not only did this
prevent a sudden message surge but it also allowed the network to
clear itself from small amounts of congestion before most endpoints
needed to send a message.
The simulation involved both network architectures with a single
application manager managing the preemption process and architectures
with multiple independent application managers. No coordination
between application managers was necessary to implement the
preemption process. Results indicated that the network was protected
from oversubscription in all cases. The only difference between
single application manager and multiple application managers occurred
at high call arrival rates where there was a small difference in the
number of sessions that were preempted for a single event. Having
multiple application managers serving the same session, with one
application manager associated with one endpoint of the sessions and
a second application manager associated with the other endpoint,
appeared to have no adverse affects on the way that the preemption
mechanism worked.
Dudley Expires - January 2006 [Page 10]
Simulation of RT-ECN based Admission Control and Preemption
5. Quantitative Performance
What follows is a summary of the results obtained through simulation.
The network used in this simulation was set up to accommodate both a
single application manager and multiple application managers. The
diagram below illustrates the distances involved and the sites
selected. At four of the sites (AK, WA, FL, DC) a bank of endpoints
was located. The link chosen to be metered was the link between KS
and WA. This selection is made not because we believe that the long
distance links are the ones likely to be congested. We actually
believe that these are the least likely to be congested. However, in
order to both get the long latencies and multiple spans of control on
a single link, it was simpler to use this link than any other.
To view the chart please see pdf version of this memo
draft-dudley-rtecn-simulation-00.pdf
All of the links used in this simulation are Gigabit Ethernet links.
The decision to do this is based partly on being able to hold all
parameters as close as possible between simulations. Overall delay
in the network is affected by both transmission delay, from the
Dudley Expires - January 2006 [Page 11]
Simulation of RT-ECN based Admission Control and Preemption
distances involved, and serialization delay, from the time it takes
to send packets of a given size. With all of the links being high
speed links, there is essentially no serialization delay. The only
delay simulated is distance delay.
The thresholds for the link being metered were engineered to perform
admission control at about 8 calls per minute. The surges applied to
the network were in the order of 240 calls per minute (4 calls per
second). Session Admission Control (SAC) illustrates admission
control or preemption cases, as noted.
5.1 Admission Control - Quantitative Analysis
To view the chart please see pdf version of this memo
draft-dudley-rtecn-simulation-00.pdf
Figure 1 Throughput Comparisons for 8 Calls per minute w/o SAC 10
Calls per Minute w/SAC
This chart illustrates the base functionality of the ECN based
Admission Control system. The first run (purple trace) is performed
at 8 calls per minute without the Session Admission Control (SAC)
scheme (8cpm_noSAC). The second run (blue trace) is performed at 10
Dudley Expires - January 2006 [Page 12]
Simulation of RT-ECN based Admission Control and Preemption
calls per minute with the Admission Control point for Routine calls
(10cppm_SAC) at 1.755 Mbps (Ethernet Throughput) and the Preemption
threshold at 2.34 Mbps (Admit, Preempt). The percentage of
precedence calls in this example is 1% for both runs.
The 10cpm_SAC run shows the traffic rising until it crosses the
admission control threshold. At this point, new routine calls are
denied. Since Precedence calls are such a small percentage of the
overall call mix, the throughput ceases to rise. At some points, the
departure of calls from the system allow it to fall back below the
threshold for a while until a new routine call takes its place.
To view the chart please see pdf version of this memo
draft-dudley-rtecn-simulation-00.pdf
Figure 2 Throughput Comparisons for 240 Calls per minute w/o SAC 240
Calls per Minute w/SAC
Dudley Expires - January 2006 [Page 13]
Simulation of RT-ECN based Admission Control and Preemption
This chart shows two runs at 240 Calls per minute. One with SAC
(240cpm_SAC) and one without (240cpm_noSAC). The Admission Control
threshold for Routine calls is 1.755 Mbps and the Preempt threshold
is 2.34 Mbps (Admit, Preempt). Both runs are at 1% Precedence.
The network is configured to have a maximum of 800 calls connected.
The run without SAC demonstrates that the number of calls rises
continuously to reach that limit. The run with SAC, however, is
limited at the Admission Control point.
5.2 Preemption - Quantitative Analysis
In order to reach the 2nd level of congestion, it was necessary to
increase the arrival rate of high precedence traffic so that it
exceeds the second level threshold. This was done by making 100% of
the traffic high precedence (called Precedence here). This is not
the normal, or expected case, so it should be noted that the results
shown here do not represent any expected behavior is a real network.
The results shown here illustrate the behavior of the system if we
choose to drive it completely with higher precedence traffic. It
doesn’t represent the expected behavior of the system but can be used
to illustrate the results of having a scenario where the admission
control point and the preemption point are at the same level. i.e.
since high precedence traffic ignores the Level 1 threshold, and
these runs show only High Precedence traffic, the admission control
and preemption control points are the same.
To view the chart please see pdf version of this memo
draft-dudley-rtecn-simulation-00.pdf
Dudley Expires - January 2006 [Page 14]
Simulation of RT-ECN based Admission Control and Preemption
To view the chart please see pdf version of this memo
draft-dudley-rtecn-simulation-00.pdf
Figure 3 Throughput Comparisons for Preemption at 10 cpm and 240 cpm
The two charts illustrate the preemption process in action to keep
the throughput (measured here in terms of numbers of sessions) at or
below the second level threshold point. The chart at 10 calls per
minute (10 cpm) clearly show the impact of having two token bucket
burst sizes, one for making the decision to start marking, and one to
make the decision to stop marking. The larger of the two burst sizes
is the decision to start marking so the excursions above the
threshold are larger than the excursions below the threshold. The
second chart appears, at times, not to make it back to the threshold.
This is, in fact, an artifact of the charting process because the
shortest time visible on the chart is a 3 second window. The maximum
value over the 3 second window was used as the reference point for
building the chart. Because of the high call arrival rate, the link
throughput only stayed below the threshold momentarily.
A second item to note form the charts is that the excursions above
the threshold are higher for the 240 cpm case than for the 10 cpm
case. This is a consequence of the latency in signaling the new
threshold level to the endpoints. Of interest is the fact that even
though the second call arrival rate is more than an order of
magnitude higher than the first rate, the size of the excursions are
Dudley Expires - January 2006 [Page 15]
Simulation of RT-ECN based Admission Control and Preemption
still small, indicating that the mechanism is quite robust to traffic
surges.
5.3 Impact of Multiple Application Managers
In our simulations, the application manager was a Back to Back User
Agent (B2BUA) that was co-located with one of the SIP endpoints in
the single application manager cases, and with each of the endpoint
sites in the 4 application manager case.. The charts that follow
compare the two cases. The reason for looking at the single
application manager and multiple application manager cases is to
examine whether the mechanism can scale to larger size networks.
To view the chart please see pdf version of this memo
draft-dudley-rtecn-simulation-00.pdf
Figure 4 Comparison of 1 and 4 B2BUA Networks at 10 Calls/Min, 1%
Precedence
The above chart illustrates the behavior of the single (1B_P01) and
four B2BUA (4B_P01) network cases when run at an arrival rate of 10
Calls/Minute. At low call arrival rates, the performance of the
single application manager and multiple application manager cases is
identical. The next chart looks at the higher call arrival rate of
240 calls/minute. It should be noted that 240 calls/minute, with a 5
minute average hold time, would result in 1200 calls on a link.
Although this would not completely fill a Gigabit Ethernet link, it
Dudley Expires - January 2006 [Page 16]
Simulation of RT-ECN based Admission Control and Preemption
is representative of the call levels that we might expect to see in a
very large network. Remembering that the thresholds are set here not
for the case where we have engineered the link for 240 calls/minute
but where the link is engineered for 8 calls/minute.
To view the chart please see pdf version of this memo
draft-dudley-rtecn-simulation-00.pdf
Figure 5 Comparison of 1 and 4 B2BUA Networks at 240 Calls/Min, 1%
Precedence
In this case, the overall appearance of the plots are similar, with
one exception on the 4B2BUA case, which dips lower on one occasion
that the 1 B2BUA case. Detailed analysis of the 4 B2BUA network
showed that each of the B2BUA is acting independently so although
each B2BUA is paced at 1 preemption per 500 milliseconds, the overall
result is that 4 preemptions will occur every 500 milliseconds if the
congestion event lasts long enough to incur multiple preemption
events.
In comparing this chart with the earlier chart illustrating the 10
call/minute scenario, we have kept the admission control and
preemption points identical and the average number of sessions is
Dudley Expires - January 2006 [Page 17]
Simulation of RT-ECN based Admission Control and Preemption
slightly higher in this case. Detailed analysis indicated that this
is caused partly by latency in the detection process. With a higher
call arrival rate, a larger number of sessions might be admitted
between the time that the threshold is actually crossed at the
network router and that the router can react and new markings arrive
at the endpoints.
6. Conclusions
This summary of test results illustrated the behaviors of an
admission control and preemption scheme based on metering link
traffic at network devices and using the ECN field to signal
congestion information to the endpoints. Taking action at the
endpoints was seen to be effective in limiting overall traffic
through the congested links. Some of the accommodations in the
metering process and the endpoint behaviors that are required to make
this system work have been noted in this summary. For admission
control, taking action at the endpoints themselves is very effective.
For preemption, the mediation of a application manager and deliberate
pacing of notifications to that server at the endpoints allows the
scheme to accommodate multiple levels of precedence. Taking action
at the endpoints allows the scheme to operate without the requirement
to track flows at network devices. The mechanics of the scheme
permit each endpoint to make independent admission control decisions.
The mechanics also permit independent preemption action to be taken
by multiple application managers so it is not necessary to track
congestion state of the entire network in a single application
manager either.
The range of behaviors possible with this general approach is quite
large. This simulation only looked at voice sessions and one set of
behaviors that could be implemented for them. Since action is taken
independently by each endpoint, it is feasible for a different set of
behaviors to be defined for different types of applications on the
same network. These different sets of behaviors could co-exist
without harm to the network as long as they were based on the same
ECN semantics and they provided a reasonable guarantee of providing
admission control and preemption limits.
Work is currently underway to investigate the performance of RT-ECN
mechanism with video conferencing systems. The variability of packet
size makes it feasible for natural variations in throughput to
occasionally reach a Level 2 threshold, which suggest that the
decision process in the endpoint for video traffic may need to be
slightly different than that needed for constant bit rate voice.
Dudley Expires - January 2006 [Page 18]
Simulation of RT-ECN based Admission Control and Preemption
Security Considerations
The studies on which this summary is based did not consider security
impacts of implementing admission control and preemption schemes.
References
i Babiarz, J. et al, “Congestion Notification Process for Real-Time
Traffic, draft-babiarz-tsvwg-rtecn-03”, Feb 2005
ii Alexander, C., “RTP Payload Format for ECN Probing, draft-
alexander-rtp-payload-for-ecn-probing”, May 2005
Intellectual Property Statement
The IETF takes no position regarding the validity or scope of any
Intellectual Property Rights or other rights that might be claimed to
pertain to the implementation or use of the technology described in
this document or the extent to which any license under such rights
might or might not be available; nor does it represent that it has
made any independent effort to identify any such rights. Information
on the procedures with respect to rights in RFC documents can be
found in BCP 78 and BCP 79.
Copies of IPR disclosures made to the IETF Secretariat and any
assurances of licenses to be made available, or the result of an
attempt made to obtain a general license or permission for the use of
such proprietary rights by implementers or users of this
specification can be obtained from the IETF on-line IPR repository at
http://www.ietf.org/ipr.
The IETF invites any interested party to bring to its attention any
copyrights, patents or patent applications, or other proprietary
rights that may cover technology that may be required to implement
this standard. Please address the information to the IETF at
ietf-ipr@ietf.org.
Disclaimer of Validity
This document and the information contained herein are provided on an
"AS IS" basis and THE CONTRIBUTOR, THE ORGANIZATION HE/SHE REPRESENTS
OR IS SPONSORED BY (IF ANY), THE INTERNET SOCIETY AND THE INTERNET
ENGINEERING TASK FORCE DISCLAIM ALL WARRANTIES, EXPRESS OR IMPLIED,
Dudley Expires - January 2006 [Page 19]
Simulation of RT-ECN based Admission Control and Preemption
INCLUDING BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF THE
INFORMATION HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED
WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.
Copyright Statement
Copyright (C) The Internet Society (2005). This document is subject
to the rights, licenses and restrictions contained in BCP 78, and
except as set forth therein, the authors retain all their rights.
Acknowledgments
Funding for the RFC Editor function is currently provided by the
Internet Society.
Author's Addresses
Stephen Dudley
Nortel
4001 E. Chapel Hill Nelson Highway
P.O. Box 13010, ms 570-01-0V8
Research Triangle Park, NC 27709
Email: SMDudley@nortel.com
Dudley Expires - January 2006 [Page 20]