Академический Документы
Профессиональный Документы
Культура Документы
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.
Evaluation
New Systems
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 dene 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 rened 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.
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
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).
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 prole.
Delay
In this dissertation, we ignore the delay introdu
ed by the playout buer. We also
assume that:
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.
PSTN
IP-core
network
PSTN
In our model, the end-to-end delay for VoIP are broken down to three omponents:
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)
Network
Modify Vat code to emulation: Send packets to a
repacketize raw bytes drop packets specified port of the
into RTP packets 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 denitions: 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
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.
Tra es olle ted from speakers who were giving a le ture or leading a dis ussion are
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
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
lassied as
onferen
e .
all type
The tra es from pre-re orded te hni al demonstrations and le ture are lassied as
Voi e tra es when two speakers were engaged in so ial onversations or te hni al
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
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
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.
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.
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 veried in our laboratory's test-bed. We will provide an overview
of the implementation and performan
e measurements in Chapter 6.
3.4 Summary