39

Chapter 3

Methodology
As des ribed in Chapter 1, the main ontribution of this thesis work is a ompre-
hensive study of a lightweight, s alable ontrol ar hite ture built on top of Di-Serv pa ket
level me hanisms to provide better end-to-end QoS support for laten y-sensitive appli a-
tions. This hapter outlines a set of resear h methodologies that we follow to arry out our
analysis, spe i ally, how we formulate the problems, examine existing systems and eval-
uate our proposed solutions. Se tion 3.1 presents the general framework that guides our
study. In Se tion 3.2, we dis uss how we model the various workloads and their performan e
requirements that later drive the performan e analysis of our ar hite ture. In Se tion 3.3,
we des ribe the evaluation methodology: we ondu ted simulation experiments using both
real tra es and generated traÆ , with a range of parameters driven by dierent performan e
goals.

3.1 General Framework

Our resear h methodology is best summarized by Figure 3.1. We follow an iterative

pro ess that omprises the following three phases:
 40

Design new architecture Analyze problem space

& mechanisms
•Model workloads
•Clearing House Analysis
Existing Systems •Define QoS requirements
•Predictive Reservations •Understand prior work
•Admission control
•Traffic policing
Design
Next
Generation

Evaluation
New Systems

Evaluate & refine solutions

• Trace-based simulation (C, Matlab)
•ns simulation

Figure 3.1: Iterative \Analysis, Design & Evaluation" phases.

 Analysis
Before we know what onstitutes a better QoS ontrol ar hite ture, we rst need to
model the hara teristi s of typi al workloads and dene their performan e require-
ments. In parti ular, we fo us on the laten y-sensitive appli ations (LSAs) su h as
pa ket audio and video be ause this type of traÆ require resour e guarantees that
are not supported by the urrent Internet. We will dis uss mathemati al models that
apture the essen e of real Internet workload. We also arry out a survey of prior
work, as reported in Chapter 2, and determine the pros and ons of ea h approa h.

 Design
In our attempt to strike a balan e between Di-Serv, whi h provides dierential treat-
ment to traÆ aggregates, and Int-Serv, whi h oers per- ow guarantees, we explore a
new ar hite ture alled the Clearing House that ombines features of the two previous
approa hes. In parti ular, we implement lightweight session-level ontrol me hanisms
su h as resour e reservations, admission ontrol and traÆ poli ing on top of a state-
 41

less Di-Serv ar hite ture. The initial design has been ontinuously rened based on
dis ussions with resear hers from two major Internet Servi e Providers that operate
nationwide ba kbone networks in the United States.

 Evaluation
The performan e evaluation of our proposed ar hite ture and algorithms are based
on a ombination of tra e-based analysis, simulation experiments, and lab prototyp-
ing. We used simulations to examine s alability issues, determine the ee t of various
design parameters on system performan e, and study the trade-os involved at var-
ious operating points. The same set of experiments are often repeated for dierent
\s enarios", where we vary the network topology, aggregate workload pattern, and
individual sour e hara teristi s.

The next se tion provides detailed dis ussions on how we model pa ket audio
appli ations that represent a typi al LSA workload. Se tion 3.3 des ribes the general simu-
lation settings but the details of ea h experiment will be presented in the subsequent three
hapters where the orresponding algorithms and performan e results are dis ussed.

3.2 Workload Modeling

In this dissertation, we onsider two basi types of workload: data appli ations
that an be sent as Best-eort traÆ , and laten y sensitive appli ations that require re-
sour e reservations and are sent as High-priority traÆ . We have hosen Voi e over IP
(VoIP) as a representative workload of the latter, be ause intera tive two-way onversa-
tions pla es a mu h more stringent delay requirements than other LSAs su h as playba k
video. Se tion 3.2.1 presents the mathemati al model that des ribes VoIP traÆ . Se -
tion 3.2.2 do uments the subje tive testing and network measurements we arry out to
quantify the performan e requirement of VoIP in terms of network entri parameters su h
 42

as delay and pa ket losses.

Besides VoIP, there are a wide variety of Internet audio appli ations that are also
laten y sensitive, in luding multimedia onferen ing, distan e learning, et . To in lude
these other appli ations in our analysis, we olle ted 70 pa ket audio tra es from te h-
ni al meetings, broad asted le tures, and multimedia onferen ing sessions. Se tion 3.2.3
do uments the olle tion and analysis of these tra es.

3.2.1 VoIP TraÆ Model

VoIP refers to real-time delivery of pa ket voi e a ross networks using the Internet
proto ols. The rapid growth of IP-based pa ket swit hed networks and the overall band-
width eÆ ien y of an integrated IP network make it an attra tive andidate to transport
voi e onne tions. In fa t, multiplexing data and voi e results in a better bandwidth utiliza-
tion than the traditional ir uit-swit hed voi e-or-nothing ba kbone in the PSTN (Publi
Swit hed Telephone Networks), whi h onsists of over-engineered voi e trunks. This justies
looking at VoIP as a workload for future Internet pa ket networks.
With silen e suppression, ea h VoIP sour e an be modeled as an on-o Markov
pro ess. The alternating periods of a tivity and silen e are exponentially distributed with
average durations of 1= and 1=, respe tively. An exponential variable X has the following
density fun tion: 8
>< ae ax x > 0; a > 0
fX (x) = >
: 0; otherwise
where E [X ℄ = 1=a and var[X ℄ = 1=a2 .
The fra tion of time that the voi e sour e is \on" is +  . We onsider an average
talk spurt of 30.83% and average silen e period of 61.47% as re ommended by the ITU-T
spe i ation [76℄ for onversational spee h. In all our experiments, we set 1= and 1=
to be 1.004 s and 1.587 s, respe tively. When the sour e is in the \on" state, xed-size
 43

pa kets are generated at a onstant interval. No pa kets are transmitted when the sour e
is \o". The size of the pa ket and the rate at whi h the pa kets are sent depends on the
orresponding voi e ode s and ompression s hemes.
Let Xi (t) be the instantaneous rate of voi e onne tion i:
8
>< R when the sour e is a tive
Xi (t) = > (3.1)
:0 when the sour e is silent

where R is the voi e bit rate (i.e., pa ket size/pa ket interval). The rate of transition from
the state of transmitting \0 Kbps" to the state of \R Kbps" is  while the reverse transition
happens at the rate of .
Traditionally voi e is Pulse Code Modulated (PCM) [77, 78℄ at 64 Kbps in the
PSTN. PCM provides high quality reprodu tion of spee h and omparable quality an be
maintained with ADPCM [79℄. Re ent advan es in ompression te hnology have allowed
highly ompressed spee h (16 Kbps and lower) that oer ex ellent voi e quality in the
absen e of pa ket losses. In our experiments, we assume that the voi e sour e generates
onstant bit rate (CBR) traÆ of 80 Kbps when it is \a tive". 1 We use this on-o Markov
pro ess to generate VoIP traÆ in our simulations (EXP1 model in Se tion 3.3).

3.2.2 VoIP Performan e Requirements

High quality intera tive voi e imposes many performan e requirements on the un-
derlying transport network. For example, one way end-to-end delay should be less than
150 ms to preserve the quality of intera tive ommuni ation. In a ir uit swit hed network,
propagation delay is the only signi ant omponent in the one way end-to-end delay. In
addition, this delay is onstant omponent during the entire all duration, and therefore an
be easily ontrolled. VoIP ar hite ture, on the other hand, introdu es new delay ompo-
1
Assume 8 KHz, 8 bits/sample PCM ode was used with 20 ms frame per pa ket. With 12 byte RTP
header, 8 byte UDP header and 20 byte IP header, the size of ea h voi e pa ket = 20 (header) + 160 (data)
= 200 bytes. The bandwidth required will be (200 x 8) bits/20 ms = 80 Kbps.
 44

nents su h as: oding/de oding delay, pa ketization delay, queuing delays at intermediate
routers/swit hes, and jitter ompensation delay introdu ed by playout buers. The mul-
tiplexing of VoIP and data traÆ on shared links also introdu es pa ket losses aused by
buer over ow at ongested nodes. Laten y and pa ket losses have adverse impa t on the
per eived voi e quality, and therefore need to be bounded.
Our goal is to show how high quality voi e an be supported with maximum utiliza-
tion of resour es if the network resour e is provisioned properly and distributed admission
ontrol is implemented. To a hieve this, we need to quantify the performan e requirements
of VoIP, by mapping the human per eived voi e quality to the more tangible network entri
parameters: pa ket loss and pa ket delay. Proper resour e provisioning te hniques an then
be applied to provide statisti al guarantees su h as upper-bound for delay or loss prole.

Delay
In this dissertation, we ignore the delay introdu ed by the playout buer. We also
assume that:

 the end-to-end propagation delay is relatively onstant and an be easily estimated,

 the sender uses the same ode throughout the all duration, and

 the sampling rate and pa ket size is xed at the beginning of ea h all.

Sin e we are interested in investigating the ee t of bandwidth allo ation on voi e

quality, we try to segregate the ee ts of appli ation-level QoS me hanisms. We assume
that no appli ation-level ongestion ontrol or rate adaptation are deployed at the voi e
sour es. The only highly variable delay omponent in our model is queuing delay that
o urs due to the multiplexing of voi e pa kets, as well as the integration of voi e and data
over a shared link.
 45

PSTN
IP-core
network

PSTN

Wireless access network

Transcoding & Propagation Delay + Queuing Delay

Packetization Delay (relatively fixed) (variable)
(negligible)

Figure 3.2: End-to-end delay omponents.

In our model, the end-to-end delay for VoIP are broken down to three omponents:

pa ketization=trans oding + propagation + queuing delay;

as shown in Figure 3.2.

ITU-T Re ommendation G. 114 [80℄ spe ies that one-way transmission time for
onne tions with adequately ontrolled e ho should be in the 0-150 ms range to be a eptable
for most user appli ations. We assume PCM trans oding introdu es almost negligible delay
if implemented in hardware (0.75 ms). The propagation delay is relatively onstant and an
be easily estimated. From [80℄, Publi Land Mobile Systems ontribute around 80 - 110 ms
to one-way propagation time. Satellite systems introdu e 12 ms at 1400 km altitude, and
110 ms at 14,000 km altitude. Opti al ber able system ontributes around 50-60 ms from
oast to oast in the United States. Assuming it takes 100 ms propagation delay for voi e
pa kets to be transported a ross the United States, the total queuing delay should be kept
within 50 ms (150ms - propagation delay). Sin e queuing delay is the only variable part in
 46

our model, we need to budget the per hop queuing delay. From about 50 tra eroute2 [81℄
experiments, we found out that there were typi ally around 8-12 hops between a ma hines
on the west oast and the east oast. Assuming that queuing delay is almost the same for
ea h hop, we require the per hop queuing delay to be at most 5 ms and use this upper-bound
to hoose appropriate buer size.

Pa ket Loss
Pa ket losses an ause further distortion beyond the unavoidable loss of infor-
mation introdu ed by spee h en oding/de oding and therefore should be minimized. We
onsider pa ket losses that are aused by buer over ows in routers as well as dis arding of
delayed pa kets in the re eiver playout buer (i.e., if pa kets arrive at the re eiver after too
long a delay and miss the playout time, these pa kets are dis arded and therefore onsidered
lost). The impa t of pa ket loss on voi e quality is dependent on the voi e ode used.
In the Fall of 1998, we used Visual Audio Tool (vat) [82℄ to run a simple subje tive
test to map the pa ket loss rate to per eived voi e quality . Vat is a multi-party audio
onferen ing tool enabled by IP-Multi ast [83℄. We onsider the following ase: PCM ode
with silen e suppression, 8 kHz sampling rate, 8 bits per sample ( ontributing to 64 Kbps
when the sour e is a tive), and 20 ms of voi e samples per pa ket.
Figure 3.3 shows the experimental setup for the subje tive test. The sound les
of three senten es (about 6 se onds ea h) from the movie, \A Few Good Men" were down-
loaded and onverted to PCM format with 8 kHz sampling rate. Sin e these sound les
are in WAV format, we used sndrfmt program to resample the voi e at 8KHz and onvert
the format to PCM and saved as -law bytes. sndrfmt 3 is a sound utility program that
2
Tra eroute an be used to display the path taken by pa kets a ross network from one host to another host.
This tool works by sending a series of UDP pa kets with dierent port numbers and TTL (Time To Live).
A list of publi servers that oer tra eroute query servi e an be found at http://www.tra eroute.org/.
3
The omplete \dpwelib" pa kage that in ludes sndrfmt and many other utility programs an be down-
loaded from http://www.i si.berkeley.edu/~dpwe/dpwelib.html.
 47

Record perceived
quality as
Sound files mean opinion score
(recorded sentences ) ( 0 -5)

Resample & Receive and

reformat sound files playback using Vat
to PCM law bytes (Listen to the
using sndrfmt specified port)

Modify Vat code to
repacketize raw bytes
into RTP packets
Network
emulation: Send packets to a
drop packets specified port of the
randomly with receiving machine
probability ploss

Figure 3.3: Experiment setup to arry out the subje tive test that maps human per eived
voi e quality to dierent pa ket loss rates.

uses a library of audio hardware and sound-le a ess fun tions developed by Dan Ellis at
the International Computer S ien e Institute (ICSI), Berkeley, CA. The voi e samples were
then pa ketized into RTP [5℄ pa kets with 12-byte RTP Header and sent through a simple
network emulation that introdu ed random pa ket losses at dierent loss rates, ploss . The
pa kets are sent at every 20 ms interval. We ran vat at the re eiving ma hine to listen to a
spe i port and playba k the data. The per eived voi e quality was s ored on a numeri
0 to 5 s ales with the following denitions: 5 = rystal lear, 4 = omprehensible but
less lear; 3 = hoppy spee h; 2 = harder to omprehend senten es due to noise; 1 = an
omprehend less than 50% of the senten e; 0 = gibberish noise. The same experiment was
repeated for dierent ploss , whi h was varied between 0 and 10%.
The result is plotted in Figure 3.4. Results show that the tolerable loss rates are
within 1-2.5% and the spee h be omes in omprehensible when more than 4% of the voi e

pa kets are lost. Note that pa ket voi e using Forward Error Corre tion (FEC) [4℄ is more
resilient to losses and therefore we would expe t the urve to shift to the right in this
ase. On the other hand, the quality of voi e onne tion using ompressed spee h is more
 48

Perceived Voice Quality

3

Satisfactory Quality

Intolerable Quality
Excellent Quality

0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Probability of Packet Loss

Figure 3.4: Subje tive test results: how pa ket loss rate ae t per eived voi e quality.

sensitive to lost voi e samples, and we expe t the urve to shift to the left. The impa t
of pa ket loss on voi e quality depends on the ode used, burstiness of losses, and frame
sizes per pa ket, but this is out of s ope of this proje t. For the rest of our analysis, we set
the upper-bound pa ket loss rate to 1%, i.e., the QoS requirement is to send high-priority
traÆ from end-to-end with at most 1% loss rate.

3.2.3 Pa ket Audio Tra es

To extend our analysis beyond VoIP, we olle ted 70 pa ket audio tra es from
a wide range of multimedia appli ations, in luding te hni al onferen e meetings, weekly
le tures, te hni al demonstrations and so ial onversations. Based on these tra es, we gen-
erate Internet workloads that have diverse hara teristi s to drive a subset of our simulation
experiments.
These tra es are lassied into the following ve ategories a ording to their audio
ontent:
 49

Table 3.1: Summary of traÆ tra es.

Type Number Duration Voi e Data
of Tra es (minutes) (pa kets, MBytes)
Audien e 32 (min) 1.26 616 pkt, 0.21 MB
(max) 123.6 3747 pkt, 1.27 MB
Classroom le ture 11 (min) 4.4 6488 pkt, 2.21 MB
(max) 71.8 100237 pkt, 34.1 MB
Conferen e all 26 (min) 0.5 528 pkt, 0.18 MB
(max) 108.2 26819 pkt, 9.12 MB
Conversation 24 (min) 1.2 1553 pkt, 0.11 MB
(max) 20.8 4287 pkt, 0.31 MB
Pre-re orded spee h 1 6.6 9781 pkt, 3.33 MB

 Tra es olle ted from speakers who were giving a le ture or leading a dis ussion are

lassied as , where the audien e (other parti ipants) may

le ture in lassroom type

interrupt the speakers with questions, leading to o asionally long pauses in the speak-
ers' voi e stream.

 Tra es that represent parti ipants who remain silent most of the time ex ept for

o asional questions or te hni al dis ussions are lassied as audien e .

type

 In a situation where the speakers were parti ipating in a group dis ussion or multi-

media onferen e alls, the tra es may have longer silen e periods su h as time spent
listening to other parti ipants or looking at a shared media board. These tra es are
lassied as onferen e .
all type

 The tra es from pre-re orded te hni al demonstrations and le ture are lassied as

pre-re orded spee h type .

 Voi e tra es when two speakers were engaged in so ial onversations or te hni al

dis ussions are lassied as onversation .

type

The traÆ tra es were generated by the following four sour es and the breakdown
of the traÆ is tabulated in Table 3.1.
 50

 CSCW Ele troni Classroom

58 tra es were olle ted from a weekly Computer S ien e graduate-level lass, Computer-
Supported Cooperative Work (CSCW)[84℄ over 14 weeks in the Fall 1997. CSCW
experimented with the idea of \Ele troni Classroom" that was well-equipped with
ollaborative te hnology su h as omputers, video ameras, monitors, and a Xerox
Live-Board. The lass was held in a small, onferen e-style room. Some students
would attend the ourse from their own oÆ e using remote ollaboration tools (e.g.,
MASH tools like vi , vat and mb). 11 tra es are lassied as lassroom le ture , 32 as
audien e, and 15 as onferen e all .

 Resear h Groups' Multimedia Conferen ing

11 tra es were re orded from onferen e alls between professors, sta members, stu-
dents and industrial sponsors of two resear h groups during January-September, 1998
and April-De ember 1999. All the tra es are lassied as onferen e all .

 Pre-re orded Te hni al Demonstrations

We in lude in our analysis voi e stream from pre-re orded te hni al demonstrations
by graduate students, whi h we lassied as pre-re orded spee h .

 CTS Test-bed with H.323 Gateway

24 tra es were re orded from a tual telephone onversations between students using
the Computer Telephony Servi e (CTS) test-bed [85℄ from January-April 2000, where
alls were made either from omputer to omputer, omputer to normal PSTN phone
or vi e-versa via a H.323 Gateway.

All the parti ipants in the CSCW lass and multimedia onferen ing ommuni-
ate through three primary kinds of media: video, audio and shared white-board, using
MASH [86℄ tools: vi , vat and \MediaBoard" (mb), respe tively. These appli ations are
laun hed on either Window-NT ma hines or Unix ma hines running Free-BSD. We are only
 51

interested in the voi e pa kets re orded in these sessions/le ture.

Tra e Pro essing

The voi e tra es were re orded a ording the MASH ar hive le formats [87, 88℄.
All data pa kets of one media type from a single sour e were stored in one le. Information
su h as the media type, the sour e identity, starting and ending time stamp were ontained
in the le header. The sender time stamp, re eiver time stamp and sequen e number of
ea h pa ket were re orded. Voi e pa kets were sent using RTP transmission format and 8
KHz 8 bits/sample PCM ode was used with 40 ms frame per pa ket. During the \talk"
state, 340 bytes pa kets were generated every 40 ms (with 12 byte RTP, 8 byte UDP header
and 320 bytes voi e data).
We determined the talkspurt and silen e periods by examining the interval be-
tween sender time stamps and lo ating gaps that were greater than 100 ms. Sin e the
smallest meaningful element of spee h, the phoneme, has an average size of 80-100 ms, we
interpreted a pause smaller than 100 ms as a stop onsonant or a minor break within the
same talkspurt. We only ran statisti al analysis on spe i segments of the voi e tra es
where a tual onversations or le ture were in progress, and the rest of the tra es were trun-
ated. For example, a speaker sometimes had to restart his/her session be ause one of the
tools (e.g., vi or \MediaBoard") failed to fun tion. Although the voi e pa kets were still
re orded from the vat session, we trun ated the pa kets re orded during the disruptions.
Se tion 3.3 will dis uss how these tra es are used in our simulation study.

3.3 Performan e Evaluation

We have designed a new ontrol ar hite ture and resour e provisioning me hanisms
to deliver better QoS support to VoIP type workload outlined in previous se tion. The
following three hapters (Chapter 4, 5, and 6) present the details of our proposed solutions
 52

and the design rationales behind them. To evaluate how well these me hanisms a hieve our
goals, we rely on a ombination of simulation study and lab prototyping using both real-
world and simulated topology. Besides network eÆ ien y and end-to-end performan e, we
explore the ar hite tural, s alability and pra ti ality issues. It is also important to identify
the degrees of freedom we have, e.g., the parameters that tune the several algorithms, and
how they ae t the trade-os among ontradi ting performan e goals, e.g., individual ow
performan e vs. overall network utilization.

3.3.1 Simulation Framework

Sin e it is infeasible to run large-s ale experiments over a tual wide-area networks,
we resort to the following simulation experiments that apture the riti al aspe ts of real-life
Internet workloads and router te hnology:

Tra e-based Simulation in C & Matlab

We developed a dis rete-time event-driven C-simulator that implements the ontrol
logi of the Clearing House ar hite ture and me hanisms. Two important inputs to the
simulator are workload models and network topology. The arrival rate of the high-priority
traÆ is modeled as an independent Poisson pro ess of intensity  alls per se ond, and
randomly pi k from the pool of 70 tra es (Se tion 3.2.3) to generate individual pa ket audio
streams. We use the topology shown in Figure 3.5, whi h is an approximation of the AT&T
WorldNet IP ba kbone as reported in [46℄.
With this simulator, we explored the eÆ ien y and robustness of the CH-ar hite ture
in terms of resour e utilization, all reje tions and reservation setup time. The details of
the experimental settings and simulation results are presented in Chapter 5. We also used
Matlab to analyze the hara teristi s our proposed reservation s heme based on Gaussian
traÆ predi tors.
 53

Boston
Chicago
Seattle
NY

St. Louis DC
Denver
SF
Atlanta

LA
Orlando
Houston

Figure 3.5: An example topology of a rst-tier IP ba kbone network in the United States.

ns Simulation
To study the pa ket-level dynami s and further evaluate our system in extreme
ases with diverse types of workload, we ran more experiments using the ns simulator [89℄.
We onstru ted an overlay network on top of ns-obje ts su h as nodes and links (imple-
mented in C++), and added session-level ontrol at the t l level. We added modules to
generate and pro ess ontrol messages transmitted using the ns UDP/IP proto ol sta k.
To evaluate the robustness of our proposed me hanisms against the diversity of
Internet workloads, we onsider four kinds of traÆ sour e models in our simulations: EXP1,
EXP2, CBR and PARETO. Ea h of these models has its own distin t statisti al properties
and an be used to represent a variety of laten y sensitive appli ations, as dis ussed in the
following.

1. EXP1 has exponential on and o times as des ribed in Se tion 3.2.1 with an average
of 1.004 s and 1.587 s, respe tively. The peak transmission rate is 80 Kbps, and the
average is approximately 31 Kbps. EXP1 an be used to model voi e appli ations,
e.g., VoIP and audio onferen ing, whi h use silen e suppression.
 54

2. EXP2 also has exponential on and o times, but with an average of 100 ms and 900
ms, respe tively. The peak rate is in reased to 310 Kbps while keeping the average
rate the same as EXP1, leading to a burstier sour e. EXP2 generates the most bursty
traÆ among the four models that we onsider and an be used to des ribe other
workloads, su h as video streams or multimedia onferen ing appli ations, that have
higher statisti al variability than VoIP.

3. CBR is a onstant bit rate sour e of 80 Kbps. Without silen e suppression, pa ket
voi e streams an be represented using CBR model.

4. PARETO sour e has Pareto on and o times and has the same peak transmission
rate (80 Kbps) as EXP1. A general Pareto density fun tion is hara terized by a
shape parameter a and a s ale parameter b:
aba
f (x) = for x  b:
xa+1
We set a = 1.5, and b is hosen su h that the on and o times have the same average as
EXP1 (1.004 s and 1.587 s, respe tively). The aggregation of Pareto sour es is known
to exhibit long range dependen ies [90, 91℄. Hen e, we use this model to des ribe
intera tive web appli ations that possess similar properties.

EXP1, EXP2 and CBR have exponential lifetimes with an average of 300s. The ow lifetimes
of PARETO sour es follow a log-normal distribution with average of 300 s.
We used ns to simulate dierent s enarios with these workloads to evaluate the
admission ontrol and mali ious ow dete tion s hemes. Results are do umented in Chap-
ter 6.

3.3.2 Lab Prototyping

We built a lab prototype of the Clearing House to evaluate ertain performan e
metri s that ould not be a urately quantied through simulations. One su h metri is the
 55

overhead of implementing the various monitoring and poli ing me hanisms an edge router.
We extended the Cli k router [24℄ to support all the traÆ poli ing and admission
ontrol fun tionalities of our ar hite ture. Using this implementation, we measured the per-
forman e overhead in urred at an edge router, e.g., the degradation of system throughput.
The urrent implementation works on Linux 2.2.16 and 2.2.17 kernels. Our ar hite ture
has been Operationally veried in our laboratory's test-bed. We will provide an overview
of the implementation and performan e measurements in Chapter 6.

3.4 Summary

This hapter gives an overview of our resear h methodology, in luding how we

model the workloads of interest and how we evaluate the performan e of our ar hite ture
through simulations and lab prototyping. The next three hapter do ument our te hni al
ontributions, design rationale and lessons learnt.