60
Uni Innsbruck Uni Innsbruck Informatik Informatik - - 1 1 IASTED PDCN 2004 Tutorial: IASTED PDCN 2004 Tutorial: Shaping the Future of Shaping the Future of Internet Congestion Control Internet Congestion Control Michael Welzl Michael Welzl http://www.welzl.at http://www.welzl.at Distributed and Parallel Systems Group Distributed and Parallel Systems Group Institute of Computer Science Institute of Computer Science University University of of Innsbruck Innsbruck

IASTED PDCN 2004 Tutorial: Shaping the Future of Internet ...home.ifi.uio.no/michawe/research/publications/iasted-pdcn2004-tuto… · Uni Innsbruck Informatik -2. Outline. 1.Congestion

  • Upload
    others

  • View
    1

  • Download
    0

Embed Size (px)

Citation preview

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    11

    IAST

    ED P

    DCN

    200

    4 Tu

    tori

    al:

    IAST

    ED P

    DCN

    200

    4 Tu

    tori

    al:

    Shap

    ing

    the

    Futu

    re o

    fSh

    apin

    g th

    e Fu

    ture

    of

    Inte

    rnet

    Con

    gest

    ion

    Cont

    rol

    Inte

    rnet

    Con

    gest

    ion

    Cont

    rol

    Mic

    hael

    Wel

    zl

    Mic

    hael

    Wel

    zl

    http

    ://w

    ww

    .wel

    zl.a

    tht

    tp:/

    /ww

    w.w

    elzl

    .at

    Dis

    trib

    uted

    and

    Par

    alle

    l Sys

    tem

    s G

    roup

    Dis

    trib

    uted

    and

    Par

    alle

    l Sys

    tem

    s G

    roup

    Inst

    itut

    e of

    Com

    pute

    r Sc

    ienc

    eIn

    stit

    ute

    of C

    ompu

    ter

    Scie

    nce

    Uni

    vers

    ity

    Uni

    vers

    ity

    ofofIn

    nsbr

    uck

    Inns

    bruc

    k

    http://www.welzl.at/

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    22

    Out

    line

    Out

    line

    1.Co

    nges

    tion

    Con

    trol

    : a

    quic

    k in

    trod

    ucti

    on

    2.Pr

    oble

    ms

    3.So

    me

    prop

    osed

    enh

    ance

    men

    ts

    4.H

    ow t

    o de

    sign

    you

    r ow

    n co

    nges

    tion

    con

    trol

    mec

    hani

    sm

    Rese

    arch

    pro

    cedu

    re:

    1.ge

    t to

    kno

    w t

    he f

    ield

    2.id

    enti

    fy p

    robl

    ems

    3.lo

    ok a

    t w

    hat

    othe

    rs

    have

    don

    e

    4.co

    me

    up w

    ith

    your

    ow

    n so

    luti

    ons

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    33

    Cong

    esti

    on C

    ontr

    olCo

    nges

    tion

    Con

    trol

    A qu

    ick

    intr

    oduc

    tion

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    44

    Prob

    lem

    sta

    tem

    ent

    Prob

    lem

    sta

    tem

    ent

    •Ef

    fici

    ent

    tran

    smis

    sion

    of

    data

    str

    eam

    s ac

    ross

    the

    Inte

    rnet

    –va

    riou

    s so

    urce

    s, v

    ario

    us d

    esti

    nati

    ons,

    var

    ious

    typ

    es o

    f st

    ream

    s

    •W

    hat

    is „

    effi

    cien

    t“?

    –te

    rms:

    late

    ncy,

    end

    2end

    del

    ay,

    jitt

    er,

    band

    wid

    th(n

    omin

    al/a

    vaila

    ble/

    bott

    lene

    ck -

    ), t

    hrou

    ghpu

    t, g

    oodp

    ut,

    loss

    rat

    io,

    ..–

    goal

    s: h

    igh

    thro

    ughp

    ut(b

    its

    / se

    cond

    ), lo

    wde

    lay,

    jit

    ter,

    loss

    rat

    io

    •N

    ote:

    Inte

    rnet

    = T

    CP/I

    P ba

    sed

    wor

    ld-w

    ide

    netw

    ork

    –no

    ass

    umpt

    ions

    abo

    ut lo

    wer

    laye

    rs!

    –ig

    nore

    CSM

    A/CD

    , CS

    MA/

    CA,

    toke

    n ri

    ng,

    base

    band

    enc

    odin

    g, f

    ram

    e ov

    erhe

    ad,

    swit

    ches

    , et

    c. e

    tc.

    !

    typically bit/s

    at this level!

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    55

    A s

    impl

    eA

    sim

    ple

    rout

    er m

    odel

    rout

    er m

    odel Q

    ueue

    1

    Que

    ue 2

    Swit

    chin

    gFa

    bric

    In 1

    Out

    1In

    2O

    ut 2

    In 3 •

    Swit

    chin

    g fa

    bric

    forw

    ards

    a pa

    cket

    (de

    st.

    addr

    .)if

    nosp

    ecia

    l tre

    atm

    ent

    nece

    ssar

    y: f

    ast

    path

    (har

    dwar

    e)

    •Q

    ueue

    sgr

    ow w

    hen

    traf

    fic

    burs

    ts a

    rriv

    e•

    low

    del

    ay=

    smal

    l que

    ues,

    low

    jit

    ter

    = m

    inor

    que

    ue f

    luct

    uati

    ons

    •Pa

    cket

    s ar

    e dr

    oppe

    dw

    hen

    queu

    es o

    verf

    low

    (“D

    ropT

    ailq

    ueue

    ing“

    )•

    low

    loss

    rat

    io=

    smal

    l que

    ues

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    66

    The

    cong

    esti

    on p

    robl

    emTh

    e co

    nges

    tion

    pro

    blem

    •Co

    nges

    tion

    cont

    rol n

    eces

    sary

    •ad

    ding

    fast

    link

    sdo

    es n

    othe

    lp!

    S1D

    1

    D2

    S2

    100

    kb/s

    1000

    kb/s

    100

    kb/s

    10kb

    /s

    110

    kb/s

    Que

    ue

    S 1

    S 2

    Swit

    chin

    gFa

    bric

    tota

    l thr

    ough

    put

    w/o

    cc.

    : 20

    kb/s

    tota

    l thr

    ough

    put

    w/

    cc.:

    110

    kb/s

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    77

    Cong

    esti

    on c

    olla

    pse

    Cong

    esti

    on c

    olla

    pse

    “knee“

    “cliff“

    Goa

    l: o

    pera

    tion

    at

    the

    “kne

    e“

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    88

    Inte

    rnet

    con

    gest

    ion

    cont

    rol:

    His

    tory

    Inte

    rnet

    con

    gest

    ion

    cont

    rol:

    His

    tory

    •19

    68/6

    9:da

    wn

    ofth

    eIn

    tern

    et•

    1986

    :fi

    rst

    cong

    esti

    onco

    llaps

    e•

    1988

    :"C

    onge

    stio

    n Av

    oida

    nce

    and

    Cont

    rol"

    (Jac

    obso

    n)Co

    mbi

    ned

    cong

    esti

    on/f

    low

    con

    trol

    for

    TCP

    •G

    oal:

    stab

    ility

    -in

    equi

    libru

    m,

    no p

    acke

    tis

    sen

    t in

    to t

    he n

    etw

    ork

    unti

    lan

    old

    pack

    etle

    aves

    –ac

    k cl

    ocki

    ng,

    “con

    serv

    atio

    nof

    pack

    ets“

    prin

    cipl

    e–

    mad

    e po

    ssib

    le t

    hrou

    gh w

    indo

    w b

    ased

    sto

    p+go

    -be

    havi

    our

    •Su

    perp

    osit

    ion

    ofst

    able

    syst

    ems

    =st

    able

    netw

    ork

    base

    don

    TCP

    wit

    h co

    nges

    tion

    con

    trol

    =st

    able

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    99

    TCP

    Cong

    esti

    on

    TCP

    Cong

    esti

    on C

    ontr

    olCo

    ntro

    l/1

    :/1

    :Ta

    hoe

    Taho

    e , 1

    988

    , 19

    88

    •D

    isti

    ngui

    sh:

    –fl

    owco

    ntro

    l:pr

    otec

    t re

    ceiv

    erag

    ains

    t ov

    erlo

    ad(r

    ecei

    ver

    "gra

    nts"

    ace

    rtai

    n am

    ount

    ofda

    ta("

    rece

    iver

    win

    dow

    ") )

    –co

    nges

    tion

    con

    trol

    :pr

    otec

    t ne

    twor

    kag

    ains

    t ov

    erlo

    ad("

    cong

    esti

    on w

    indo

    w"

    (cw

    nd)

    limit

    sth

    e ra

    te:

    min

    (cw

    nd,r

    wnd

    ) us

    ed!

    )

    •Fl

    ow/C

    onge

    stio

    nCo

    ntro

    l com

    bine

    din

    TCP

    .Se

    vera

    l alg

    orit

    hms:

    •(w

    indo

    w u

    nit:

    SM

    SS=

    Send

    er M

    axim

    um S

    egm

    ent

    Size

    , us

    ually

    adju

    sted

    to

    Path

    MTU

    ; in

    it c

    wnd

    =

    ssth

    resh

    –Co

    nges

    tion

    Avoi

    danc

    e:ea

    chRT

    T,in

    crea

    se c

    wnd

    by

    SMSS

    *SM

    SS/c

    wnd

    (lin

    ear

    grow

    th -

    "add

    itiv

    e in

    crea

    se")

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    1010

    TCP

    Cong

    esti

    onTC

    P Co

    nges

    tion

    Cont

    rol

    Cont

    rol

    /2/2

    •If

    a pa

    cket

    or a

    ck is

    lost

    (t

    imeo

    ut,

    roug

    hly

    4*rt

    t),

    set

    cwnd

    = 1,

    ssth

    resh

    =cu

    rren

    t ba

    ndw

    idth

    / 2

    (“m

    ulti

    plic

    ativ

    e de

    crea

    se")

    -ex

    pone

    ntia

    lbac

    koff

    •Se

    vera

    l tim

    ers,

    base

    don

    RT

    T; g

    ood

    esti

    mat

    ion

    is

    cruc

    ial!

    •La

    ter

    addi

    tion

    s:(T

    CP R

    eno,

    199

    0)Fa

    stre

    tran

    smit

    / fa

    stre

    cove

    ry(n

    otif

    y se

    nder

    oflo

    ss v

    ia d

    uplic

    ate

    acks

    )

    0123456789

    12

    34

    56

    78

    910

    1112

    1314

    1516

    time

    bandwidth

    Timeo

    ut

    ssth

    resh Sl

    ow Start

    Congestion

    Avoidance

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    1111

    Use

    r 1 A

    lloca

    tion

    x1

    Fairn

    ess

    Line

    Effic

    ienc

    yLi

    ne

    User 2 Allocation x2

    Sta

    rting

    Poi

    nt

    AIM

    D

    Des

    irabl

    e

    Sta

    rting

    Poi

    nt

    AIAD

    MIM

    D Und

    erlo

    ad

    Ove

    rload

    Use

    r 1 A

    lloca

    tion

    x1

    Fairn

    ess

    Line

    Effic

    ienc

    yLi

    ne

    User 2 Allocation x2

    Sta

    rting

    Poi

    nt

    AIM

    D

    Des

    irabl

    e

    Sta

    rting

    Poi

    nt

    AIAD

    MIM

    D Und

    erlo

    ad

    Ove

    rload

    Back

    grou

    nd:

    AIM

    DBa

    ckgr

    ound

    : A

    IMD

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    1212

    Act

    ive

    Act

    ive

    Que

    ue M

    anag

    emen

    tQ

    ueue

    Man

    agem

    ent

    •To

    day,

    TCP

    beha

    viou

    r do

    min

    ates

    the

    Inte

    rnet

    (W

    WW

    , ..

    )

    •(s

    omew

    hat

    old)

    exa

    mpl

    e ba

    ckbo

    ne m

    easu

    rem

    ent:

    98%

    TCP

    traf

    fic

    •19

    93:

    Rand

    om E

    arly

    Det

    ecti

    on("

    Dis

    card

    ", "

    Dro

    p")

    (RED

    )(n

    ow t

    hat

    end

    node

    sba

    ck o

    ffas

    pac

    kets

    are

    dro

    pped

    ,dr

    oppa

    cket

    s ea

    rlie

    rto

    avoi

    dqu

    eue

    over

    flow

    s)

    •An

    othe

    r go

    al:

    add

    rand

    omiz

    atio

    nto

    avoi

    d tr

    affi

    c ph

    ase

    effe

    cts!

    •Q

    avg

    = (1

    -W

    q) x

    Qav

    g+

    Qin

    stx

    Wq

    (Qav

    g=

    aver

    age

    occu

    panc

    y,Q

    inst

    =in

    stan

    tane

    ous

    occu

    panc

    y,W

    q=

    wei

    ght

    -ha

    rdto

    tun

    e,de

    term

    ines

    how

    aggr

    essi

    ve R

    EDbe

    have

    s)

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    1313

    Act

    ive

    Act

    ive

    Que

    ue M

    anag

    emen

    t /2

    Que

    ue M

    anag

    emen

    t /2

    •Ba

    sed

    onex

    pone

    ntia

    lly w

    eigh

    ted

    mov

    ing

    aver

    age

    (EW

    MA)

    of

    inst

    anta

    neou

    s qu

    eue

    occu

    panc

    y=

    low

    pass

    filt

    er–

    reca

    lcul

    ated

    eve

    ryti

    me

    a pa

    cket

    arri

    ves

    •Q

    avg

    belo

    w t

    hres

    hold

    min

    _th:

    Not

    hing

    hap

    pens

    •Q

    avg

    abov

    e th

    resh

    old

    min

    _th:

    Dro

    ppr

    obab

    ility

    ris

    es li

    near

    ly•

    Qav

    gab

    ove

    thre

    shol

    d m

    ax_t

    h:D

    rop

    pack

    ets

    •RE

    Dex

    pect

    sal

    lflo

    ws

    tobe

    have

    like

    TCP

    -bu

    t is

    itfa

    ir?

    •Va

    rian

    ts:

    drop

    from

    fron

    t, d

    rop

    base

    don

    inst

    anta

    neou

    s qu

    eue

    occu

    panc

    y, d

    rop

    arbi

    trar

    y pa

    cket

    s, d

    rop

    base

    don

    prio

    riti

    es..

    .

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    1414

    Expl

    icit

    Expl

    icit

    Cong

    esti

    onCo

    nges

    tion

    Not

    ific

    atio

    nN

    otif

    icat

    ion

    (ECN

    )(E

    CN)

    •19

    99:

    Expl

    icit

    Cong

    esti

    onN

    otif

    icat

    ion

    (ECN

    )In

    stea

    dof

    drop

    ping

    ,se

    ta

    bit

    •En

    d sy

    stem

    sar

    e ex

    pect

    edto

    act

    as if

    pack

    et w

    asdr

    oppe

    d⇒

    actu

    al c

    omm

    unic

    atio

    n be

    twee

    nen

    dno

    des

    and

    the

    netw

    ork!

    •AT

    M a

    nd F

    ram

    eRe

    lay:

    not

    only

    ECN

    but

    also

    BEC

    N

    •In

    tern

    et B

    ECN

    :of

    ten

    prop

    osed

    and

    regu

    larl

    y di

    scus

    sed

    (ICM

    P SQ

    ),bu

    t ve

    ry u

    nlik

    ely

    -se

    vera

    l rea

    sons

    •Ve

    ry p

    opul

    ar a

    mon

    g re

    sear

    cher

    s -

    lots

    of

    idea

    s to

    exp

    loit

    the

    bit

    !

    •EC

    Nca

    nnot

    tot

    ally

    rep

    lace

    loss

    mea

    sure

    men

    ts!

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    1515

    Prob

    lem

    sPr

    oble

    ms

    ( =

    pote

    ntia

    l res

    earc

    h to

    pics

    )

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    1616

    TCP

    TCP

    in h

    eter

    ogen

    eous

    env

    iron

    men

    tsin

    het

    erog

    eneo

    us e

    nvir

    onm

    ents

    •TC

    P ov

    er n

    oisy

    link

    s: p

    robl

    ems

    wit

    h „p

    acke

    t lo

    ss =

    con

    gest

    ion“

    –w

    asba

    did

    eain

    tim

    esof

    erro

    r-pr

    one

    netw

    orks

    –se

    ems

    reas

    onab

    lein

    tim

    esof

    fibr

    e ne

    twor

    ks–

    real

    ly b

    ad f

    orw

    irel

    ess

    links

    !

    •TC

    Pov

    er“l

    ong

    fat

    pipe

    s“:

    larg

    e ba

    ndw

    idth

    *del

    ay p

    rodu

    ct–

    long

    tim

    e to

    reac

    h eq

    uilib

    rium

    , M

    D =

    prob

    lem

    atic

    !

    •TC

    P in

    hig

    hly

    asym

    met

    ric

    netw

    orks

    :(e

    .g.

    dire

    ct s

    atel

    lite

    last

    mile

    )–

    inco

    min

    g th

    roug

    hput

    (hi

    gh c

    apac

    ity

    link)

    lim

    ited

    by

    rate

    of

    outg

    oing

    ACK

    s(„

    ACK

    com

    pres

    sion

    “)

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    1717

    Fair

    ness

    Fair

    ness

    •AT

    M A

    BR:

    Max

    -Min

    -fai

    rnes

    s–

    “A (

    ..)

    allo

    cati

    onof

    rate

    s is

    max

    -min

    fai

    rif

    fan

    incr

    ease

    ofan

    yra

    te (

    ..)

    mus

    t be

    atth

    e co

    stof

    ade

    crea

    seof

    som

    e al

    read

    y sm

    alle

    rra

    te.“

    –O

    nere

    sour

    ce:

    mat

    hem

    atic

    al d

    efin

    itio

    n sa

    tisf

    ies

    "gen

    eral

    "un

    ders

    tand

    ing

    offa

    irne

    ss-

    reso

    urce

    is d

    ivid

    ed e

    qual

    ly a

    mon

    g co

    mpe

    tito

    rs–

    Usu

    ally

    requ

    ires

    kno

    wle

    dge

    offl

    ows

    inro

    uter

    s(s

    wit

    ches

    ) -

    scal

    abili

    ty p

    robl

    em!

    •In

    tern

    et:

    –TC

    P do

    min

    ant,

    but

    does

    not

    sat

    isfy

    Max

    -Min

    -fai

    rnes

    s cr

    iter

    ion!

    –Ac

    k-cl

    ocke

    d-

    flow

    s w

    ith

    shor

    ter

    RTT

    reac

    t so

    oner

    (slo

    wst

    art,

    ..)

    and

    achi

    eve

    bett

    er r

    esul

    ts–

    Ther

    efor

    e, In

    tern

    etde

    fini

    tion

    offa

    irne

    ss:

    TCP-

    frie

    ndlin

    ess

    "Afl

    ow is

    TCP-

    com

    pati

    ble

    (TCP

    -fri

    endl

    y)if

    , in

    stea

    dy s

    tate

    ,it

    use

    sno

    mor

    e ba

    ndw

    idth

    tha

    na

    conf

    orm

    ant

    TCP

    runn

    ing

    unde

    r co

    mpa

    rabl

    e co

    ndit

    ions

    ."

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    1818

    Issu

    es w

    ith

    TCP

    Issu

    es w

    ith

    TCP

    -- fri

    endl

    ines

    sfr

    iend

    lines

    s

    •TC

    P re

    gula

    rly

    incr

    ease

    s th

    e qu

    eue

    leng

    than

    d ca

    uses

    loss

    ⇒de

    tect

    con

    gest

    ion

    whe

    n it

    is a

    lrea

    dy (

    ECN

    : al

    mos

    t) t

    oo la

    te!

    –po

    ssib

    le t

    o ha

    vem

    ore

    thro

    ughp

    ut w

    ith

    smal

    ler

    queu

    es a

    nd le

    ss lo

    ss..

    . bu

    t: e

    xcee

    d ra

    te o

    f TC

    P un

    der

    sim

    ilar

    cond

    itio

    ns ⇒

    not

    TCP-

    frie

    ndly

    !

    •W

    hat

    if I

    send

    mor

    e th

    an T

    CP i

    n th

    e ab

    senc

    eof

    com

    peti

    ng T

    CP‘s

    ?–

    can

    such

    a m

    echa

    nism

    exi

    st?

    –ye

    s! T

    CP it

    self

    , w

    ith

    max

    . w

    indo

    w s

    ize

    = ba

    ndw

    idth

    * R

    TT–

    Doe

    s th

    is m

    ean

    that

    TC

    P is

    not

    TCP

    -fri

    endl

    y?

    •D

    etai

    ls m

    issi

    ng f

    rom

    the

    def

    init

    ion:

    –pa

    ram

    eter

    s +

    vers

    ion

    of „

    conf

    orm

    ant

    TCP“

    –du

    rati

    on!

    shor

    t TC

    P fl

    ows

    are

    diff

    eren

    t th

    an lo

    ng o

    nes

    •TC

    P-fr

    iend

    lines

    s =

    com

    pati

    bilit

    y of

    new

    mec

    hani

    sms

    wit

    h ol

    d m

    echa

    nism

    –th

    ere

    was

    res

    earc

    h si

    nce

    the

    80‘s

    ! e.

    g. n

    ew k

    now

    ledg

    e ab

    out

    netw

    ork

    mea

    sure

    men

    ts

    •TC

    P ra

    te d

    epen

    ds o

    n RT

    T -

    how

    doe

    s th

    is r

    elat

    e to

    „fa

    irne

    ss“?

    Does TCP-friendliness

    hinder research?

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    1919

    Prop

    orti

    onal

    Fai

    rnes

    sPr

    opor

    tion

    al F

    airn

    ess

    Prop

    orti

    onal

    fair

    ness

    :“A

    nal

    loca

    tion

    ofra

    tes

    xis

    pro

    port

    iona

    llyfa

    irif

    f,fo

    r an

    y ot

    her

    (..)

    allo

    cati

    on

    y,w

    e ha

    ve:

    (rou

    ghly

    appr

    oxim

    ated

    by

    AIM

    D!)

    01

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    2020

    Cong

    esti

    on p

    rici

    ngCo

    nges

    tion

    pri

    cing

    •Ba

    sic

    idea

    :hi

    gher

    sen

    ding

    rat

    e =

    mor

    e co

    nges

    tion

    = h

    ighe

    r pr

    ice

    –Id

    ea:

    char

    ge m

    ore

    whe

    n EC

    Nfl

    ag is

    set

    •„S

    mar

    t“ M

    arke

    tid

    ea:

    –ea

    ch p

    acke

    t bi

    ds f

    or c

    apac

    ity

    –ou

    t of

    nbi

    ds,

    mhi

    ghes

    t bi

    ds c

    an b

    e ac

    cept

    ed–

    pric

    e: „

    mar

    gina

    l cos

    t“ (

    high

    est

    bid

    of u

    nacc

    epte

    d pa

    cket

    s)⇒

    lead

    s to

    mar

    ket

    equi

    libri

    um

    •„S

    mar

    t“ M

    arke

    t no

    t pr

    acti

    cal (

    bidd

    ing

    per

    pack

    et),

    but

    sho

    ws:

    –ba

    lanc

    e of

    dem

    and

    and

    supp

    ly le

    ads

    to m

    arke

    t eq

    uilib

    rium

    –st

    able

    sys

    tem

    jus

    t ba

    sed

    on e

    cono

    mic

    s su

    ffic

    es f

    or c

    onge

    stio

    n co

    ntro

    l

    •Pr

    oble

    m:

    link

    cann

    ot k

    now

    bid

    s in

    adv

    ance

    ⇒no

    QoS

    guar

    ante

    es

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    2121

    End2

    end

    real

    End2

    end

    real

    -- tim

    eti

    me

    data

    tra

    nsfe

    rda

    ta t

    rans

    fer

    •As

    sum

    ptio

    n: n

    osp

    ecia

    l ser

    vice

    ava

    ilabl

    eat

    appl

    icat

    ion

    leve

    l–

    (Def

    init

    ion

    of In

    tern

    et "

    real

    -tim

    e"so

    fter

    than

    usu

    al)

    •D

    iffe

    rent

    requ

    irem

    ents

    :–

    relia

    ble

    serv

    ice

    may

    not

    be

    need

    ed(n

    ore

    tran

    smis

    sion

    )–

    Tim

    ely

    tran

    smis

    sion

    impo

    rtan

    t

    –D

    iffe

    rent

    trea

    tmen

    t:–

    nore

    tran

    smis

    sion

    / w

    aiti

    ng f

    or A

    CKs

    –no

    slid

    ing

    win

    dow

    (sto

    p +

    go b

    ehav

    iour

    not

    sui

    tabl

    e)–

    but:

    –so

    me

    kind

    offl

    ow c

    ontr

    olst

    illne

    eded

    –sy

    nchr

    oniz

    atio

    n ne

    cess

    ary

    –of

    ten:

    Mul

    tica

    st

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    2222

    Mul

    tim

    edia

    ada

    ptat

    ion

    Mul

    tim

    edia

    ada

    ptat

    ion

    •M

    ista

    ke:

    –ad

    apta

    tion

    sch

    emes

    ass

    ume

    arbi

    trar

    y da

    ta s

    trea

    m s

    cala

    bilit

    y

    •Pr

    oble

    m:

    –D

    ata

    stre

    ams

    show

    flu

    ctua

    tion

    s (e

    xam

    ple:

    MPE

    G I-

    , B-

    , P-

    fram

    es)

    •So

    luti

    on:

    –Sp

    ecia

    l CBR

    des

    ign

    for

    com

    mun

    icat

    ion

    -H

    .261

    des

    igne

    d fo

    r IS

    DN

    –no

    t al

    way

    s fe

    asib

    le

    •Pr

    oble

    m:

    –co

    mpr

    essi

    on u

    sual

    ly n

    ot d

    eter

    min

    isti

    c -

    size

    dep

    ends

    on

    cont

    ent!

    –re

    al-l

    ife

    dist

    ance

    lear

    ning

    exa

    mpl

    e:40

    kbps

    eno

    ugh

    for

    stre

    amin

    g vi

    deo

    (Sm

    artb

    oard

    ) +

    audi

    o (s

    peec

    h),

    but

    spee

    ch s

    uffe

    rs d

    ram

    atic

    ally

    if t

    each

    er v

    isib

    le

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    2323

    Cong

    esti

    on C

    ontr

    ol a

    nd Q

    ualit

    y of

    Ser

    vice

    Cong

    esti

    on C

    ontr

    ol a

    nd Q

    ualit

    y of

    Ser

    vice

    •Q

    ualit

    y of

    Ser

    vice

    (Q

    oS):

    prov

    ide

    diff

    eren

    tiat

    ed q

    ualit

    y ba

    sed

    on $

    $$

    •Q

    uest

    ions

    :–

    is a

    flu

    id lo

    w-q

    ualit

    y vi

    deo

    bett

    er t

    han

    a bu

    ckin

    g hi

    gh-q

    ualit

    y vi

    deo?

    –an

    d w

    hat

    abou

    t au

    dio?

    •Sc

    alab

    le Q

    oS =

    no

    per-

    flow

    gua

    rant

    ees

    –st

    anda

    rd a

    rchi

    tect

    ure:

    Dif

    fere

    ntia

    ted

    Serv

    ices

    (D

    iffS

    erv)

    plac

    es u

    sers

    in f

    low

    agg

    rega

    tes

    –co

    nges

    tion

    con

    trol

    sti

    ll ne

    cess

    ary

    wit

    hin

    an a

    ggre

    gate

    ⇒pe

    r-fl

    ow Q

    oS d

    epen

    ds o

    n co

    nges

    tion

    con

    trol

    mec

    hani

    sm!

    –H

    ow d

    oes

    a co

    nges

    tion

    con

    trol

    mec

    hani

    sm in

    tera

    ct w

    ith

    QoS

    ele

    men

    ts?

    e.g.

    TCP

    kno

    wn

    to b

    e a

    bad

    mat

    ch f

    or t

    oken

    buc

    ket

    ...modest QoS with adaptive

    multimedia? -

    still unresolved!

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    2424

    Spec

    ial t

    ypes

    of

    traf

    fic

    Spec

    ial t

    ypes

    of

    traf

    fic

    •G

    rid:

    pre

    dict

    able

    tra

    ffic

    pat

    tern

    –Th

    is is

    tot

    ally

    new

    to

    the

    Inte

    rnet

    !•

    Web

    : us

    ers

    crea

    te t

    raff

    ic•

    FTP

    dow

    nloa

    d: s

    tart

    s ..

    . en

    ds•

    Stre

    amin

    g vi

    deo:

    eit

    her

    CBR

    or d

    epen

    ds o

    n co

    nten

    t! (

    head

    mov

    emen

    t, .

    .)

    •Sp

    ecia

    l req

    uire

    men

    ts a

    nd p

    rope

    rtie

    s–

    may

    req

    uire

    del

    ay b

    ound

    s or

    min

    imum

    ban

    dwid

    th–

    mix

    ture

    of

    spor

    adic

    (RP

    C ty

    pe)

    mes

    sage

    s an

    d bu

    lk d

    ata

    tran

    sfer

    •Re

    late

    d: s

    igna

    ling

    traf

    fic

    –us

    ually

    not

    a la

    rge

    amou

    nt o

    f da

    ta–

    to d

    ate,

    no

    seri

    ous

    effo

    rts

    for

    tailo

    red

    cong

    esti

    on c

    ontr

    ol(S

    CTP

    = tr

    ansp

    ort

    prot

    ocol

    des

    igne

    d fo

    r si

    gnal

    ing;

    con

    gest

    ion

    cont

    rol=

    si

    mila

    r to

    TCP

    )

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    2525

    Som

    e re

    ason

    s fo

    r TC

    P st

    abili

    tySo

    me

    reas

    ons

    for

    TCP

    stab

    ility

    “Con

    gest

    ion

    Avoi

    danc

    e an

    d Co

    ntro

    l“,

    Van

    Jaco

    bson

    , SI

    GCO

    MM

    ‘88:

    •Ex

    pone

    ntia

    l bac

    koff

    :“F

    or a

    tra

    nspo

    rt e

    ndpo

    int

    embe

    dded

    in a

    net

    wor

    k of

    unk

    now

    n to

    polo

    gy a

    nd w

    ith

    an u

    nkno

    wn,

    unk

    now

    able

    and

    con

    stan

    tly

    chan

    ging

    popu

    lati

    on o

    f co

    mpe

    ting

    con

    vers

    atio

    ns,

    only

    one

    sch

    eme

    has

    any

    hope

    of

    wor

    king

    -ex

    pone

    ntia

    l bac

    koff

    -bu

    t a

    proo

    f of

    thi

    s is

    bey

    ond

    the

    scop

    e of

    thi

    s pa

    per.

    •Co

    nser

    vati

    on o

    f pa

    cket

    s:“T

    he p

    hysi

    cs o

    f fl

    ow p

    redi

    cts

    that

    sys

    tem

    s w

    ith

    this

    pro

    pert

    y sh

    ould

    be

    rob

    ust

    in t

    he f

    ace

    of c

    onge

    stio

    n.“

    •Ad

    diti

    ve In

    crea

    se,

    Mul

    tipl

    icat

    ive

    Dec

    reas

    e:N

    ot e

    xplic

    itel

    y ci

    ted

    as a

    sta

    bilit

    y re

    ason

    in t

    he p

    aper

    !–

    ...b

    ut in

    100

    0‘s

    of o

    ther

    pap

    ers!

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    2626

    “Pro

    ofs“

    of

    TCP

    stab

    ility

    “Pro

    ofs“

    of

    TCP

    stab

    ility

    •AI

    MD

    :Ch

    iu/J

    ain:

    dia

    gram

    + a

    lgeb

    raic

    pro

    of o

    f ho

    mog

    eneo

    us R

    TT c

    ase

    •st

    eady

    -sta

    te T

    CP m

    odel

    : w

    indo

    w s

    ize

    ~ 1/

    sqrt

    (p)

    (p =

    pac

    ket

    loss

    )

    •Jo

    hari

    /Tan

    , M

    asso

    ulié

    , ..

    :–

    loca

    lsta

    bilit

    y, n

    egle

    ct d

    etai

    ls o

    f TC

    P be

    havi

    our

    (flu

    id f

    low

    mod

    el,

    ..)

    –as

    sum

    ptio

    n:“q

    ueue

    ing

    dela

    ys w

    ill e

    vent

    ually

    bec

    ome

    smal

    l rel

    ativ

    e to

    pro

    paga

    tion

    del

    ays“

    •St

    even

    Low

    :–

    Dua

    lity

    mod

    el (

    base

    d on

    uti

    lity

    func

    tion

    / F

    . Ke

    lly,

    ..):

    Stab

    ility

    dep

    ends

    on

    dela

    y, c

    apac

    ity,

    load

    and

    AQ

    M !

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    2727

    Uni

    cast

    / B

    road

    cast

    / (

    over

    lay)

    Mul

    tica

    stU

    nica

    st /

    Bro

    adca

    st /

    (ov

    erla

    y) M

    ulti

    cast

    1 Sender

    2 Receivers

    SRR

    Unicast

    SRR

    Broadcast

    SRR

    Overlay Multicast

    SRR

    IP Multicast

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    2828

    Mul

    tica

    st is

    sues

    Mul

    tica

    st is

    sues

    •Re

    quir

    ed f

    or a

    pplic

    atio

    ns w

    ith

    mul

    tipl

    e re

    ceiv

    ers

    only

    –vi

    deo

    conf

    eren

    ces,

    rea

    l-ti

    me

    data

    str

    eam

    tra

    nsm

    issi

    on,

    ..⇒

    diff

    eren

    t da

    ta s

    trea

    ms

    than

    web

    sur

    fing

    , ft

    p do

    wnl

    oads

    etc

    !

    •Is

    sues

    :–

    grou

    p m

    anag

    emen

    t•

    prot

    ocol

    req

    uire

    d to

    joi

    n /

    leav

    e gr

    oup

    dyna

    mic

    ally

    :In

    tern

    et G

    roup

    Man

    agem

    ent

    Prot

    ocol

    (IG

    MP)

    •st

    ate

    in r

    oute

    rs:

    hard

    / s

    oft

    (los

    t un

    less

    ref

    resh

    ed)?

    •w

    ho in

    itia

    tes

    / co

    ntro

    ls g

    roup

    mem

    bers

    hip?

    –co

    nges

    tion

    con

    trol

    •sc

    alab

    ility

    (ACK

    impl

    osio

    n)•

    deal

    ing

    wit

    hhe

    tero

    gene

    ity

    of r

    ecei

    ver

    grou

    ps•

    fair

    ness

    •M

    ulti

    cast

    con

    gest

    ion

    cont

    rol m

    echa

    nism

    cla

    ssif

    icat

    ion:

    –se

    nder

    -vs

    . re

    ceiv

    er-b

    ased

    , si

    ngle

    -rat

    e vs

    . m

    ulti

    -rat

    e (l

    ayer

    ed),

    –re

    liabl

    e vs

    . un

    relia

    ble,

    end

    -to-

    end

    vs.

    netw

    ork-

    supp

    orte

    d

    depends on

    content!

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    2929

    Som

    e pr

    opos

    ed e

    nhan

    cem

    ents

    Som

    e pr

    opos

    ed e

    nhan

    cem

    ents

    Rese

    arch

    by

    othe

    rs

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    3030

    Inte

    rnet

    Inte

    rnet

    traf

    fic

    char

    acte

    rist

    ics

    traf

    fic

    char

    acte

    rist

    ics

    •M

    RTG

    trac

    e(b

    ased

    on S

    NM

    P,ac

    cess

    ing

    traf

    fic

    coun

    ters

    in M

    IB)

    http

    ://w

    ww

    .switc

    h.ch

    /lan/

    stat

    /pee

    rings

    /link

    eune

    t.htm

    l, 11

    . 10.

    99,

    13:

    05

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    3131

    Inte

    rnet

    Inte

    rnet

    traf

    fic

    char

    acte

    rist

    ics

    traf

    fic

    char

    acte

    rist

    ics

    /2/2

    •Tr

    adit

    iona

    ltra

    ffic

    mod

    ellin

    g:qu

    euin

    g th

    eory

    noti

    on:

    traf

    fic

    follo

    ws

    pois

    son

    dist

    ribu

    tion

    •In

    tern

    ettr

    affi

    c is

    bur

    sty

    -in

    tuit

    ive

    reas

    ons:

    –TC

    Pis

    bur

    sty

    byna

    ture

    :co

    nges

    tion

    avoi

    danc

    e,pa

    yloa

    d vs

    .ac

    ks…

    –AC

    Kco

    mpr

    essi

    on c

    anca

    use

    payl

    oad

    burs

    ts d

    ueto

    ACK

    -clo

    ckin

    g (l

    ater

    !)

    –va

    riou

    spa

    cket

    size

    s

    –Bu

    rsts

    fro

    m q

    ueue

    s ag

    greg

    ate

    as t

    raff

    ictr

    aver

    ses

    the

    net

    –Bu

    rsti

    ness

    ofon

    e fl

    ow a

    ffec

    ts o

    ther

    adap

    tive

    flow

    s

    Still true

    for user

    arrival !

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    3232

    Inte

    rnet

    Inte

    rnet

    traf

    fic

    char

    acte

    rist

    ics

    traf

    fic

    char

    acte

    rist

    ics

    /3/3

    •O

    verl

    appi

    ngof

    inde

    pend

    ent

    on-o

    ffso

    urce

    s le

    ads

    todi

    stri

    buti

    on w

    ith

    heav

    y-ta

    iled

    auto

    corr

    elat

    ion

    func

    tion

    •Lo

    ng-r

    ange

    dep

    enda

    nce:

    "pe

    aks

    sit

    onri

    pple

    sw

    hich

    sit

    onw

    aves

    "

    •N

    o "f

    latt

    enin

    g"to

    war

    dsa

    mea

    n as

    you

    zoom

    out

    -sa

    me

    stru

    ctur

    es

    may

    be

    foun

    dat

    dif

    fere

    nt t

    ime

    scal

    es,

    henc

    e se

    lf s

    imila

    r

    •ch

    arac

    teri

    stic

    sm

    odel

    edw

    ith

    tim

    ese

    ries

    (fAR

    IMA

    mod

    els)

    or w

    avel

    ets

    –M

    easu

    rem

    ent

    ofth

    e"d

    egre

    eof

    self

    sim

    ilari

    ty":

    Hur

    stpa

    ram

    eter

    –⇒

    mod

    elap

    prox

    imat

    ion

    invo

    lves

    Hur

    stpa

    ram

    eter

    est

    imat

    ion

    •TC

    P kn

    own

    to p

    ropa

    gate

    bot

    tlen

    eck

    self

    -sim

    ilari

    tyto

    end

    sys

    tem

    –po

    ssib

    ility

    : us

    e m

    odel

    to

    pred

    ict

    traf

    fic

    inst

    ead

    of g

    uess

    ing

    –qu

    esti

    on:

    scal

    abili

    ty(w

    hat

    if e

    very

    body

    doe

    s th

    is?)

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    3333

    How

    How

    totobebe

    TCP

    TCP

    -- fri

    endl

    yfr

    iend

    ly

    •TC

    P-fr

    iend

    lines

    s ca

    n be

    ach

    ieve

    d by

    em

    ulat

    ing

    the

    beha

    viou

    rof

    TCP

    (or

    the

    des

    ired

    par

    tsof

    it)

    •Si

    mpl

    ifie

    dTC

    P: A

    IMD

    (add

    itiv

    e in

    cr. α

    ,m

    ulti

    plic

    ativ

    e de

    cr. β)

    –0

    < α

    , 0

    < β

    < 1

    ->st

    able

    and

    fair

    cong

    esti

    on c

    ontr

    ol–α

    = 4

    x (1

    -β^

    2) /

    3->

    TCP

    -fri

    endl

    y co

    nges

    tion

    con

    trol

    (G

    AIM

    D)

    –α

    = 1,

    β

    = 1/

    2->

    TCP

    •AI

    MD

    mec

    hani

    sms

    for

    mul

    tim

    edia

    app

    licat

    ions

    : RA

    P, L

    DA+

    •D

    iffe

    rent

    appr

    oach

    es:

    –TC

    PEm

    ulat

    ion

    At R

    ecei

    vers

    (TE

    AR)

    TCP

    calc

    ulat

    ions

    (cw

    nd c

    alcu

    lati

    on,

    fast

    reco

    very

    , ..

    .)m

    oved

    tore

    ceiv

    er,

    dono

    t ac

    k ev

    ery

    pack

    et,

    smoo

    th s

    endi

    ngra

    te–

    Bino

    mia

    l con

    gest

    ion

    cont

    rol:

    gene

    raliz

    atio

    n of

    GAI

    MD

    wit

    h no

    nlin

    ear

    cont

    rol

    –CY

    RF f

    ram

    ewor

    k:ge

    nera

    lizat

    ion

    of b

    inom

    ial c

    onge

    stio

    n co

    ntro

    l

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    3434

    GA

    IMD

    GA

    IMD

    cong

    esti

    on c

    ontr

    olco

    nges

    tion

    con

    trol

    Rela

    tion

    ship

    bet

    wee

    and β

    for

    TCP-

    frie

    ndlin

    ess:

    00,

    20,

    40,

    60,

    811,

    21,

    4

    00,

    10,

    20,

    30,

    40,

    50,

    60,

    70,

    80,

    91

    beta

    alpha

    TCP

    smoo

    ther

    mor

    eag

    gres

    sive

    +re

    spon

    sive

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    3535

    Equa

    tion

    bas

    ed c

    onge

    stio

    n co

    ntro

    lEq

    uati

    on b

    ased

    con

    gest

    ion

    cont

    rol

    •Ba

    sed

    on T

    CPst

    eady

    -sta

    te r

    espo

    nse

    func

    tion

    -gi

    ves

    uppe

    r bo

    und

    for

    tran

    smis

    sion

    rate

    T (

    byte

    s/se

    c):

    )32

    1()

    833(

    322 p

    pp

    tp

    R

    sT

    RTO

    ++

    =

    )32

    1()

    833(

    322 p

    pp

    tp

    R

    sT

    RTO

    ++

    =

    s: p

    acke

    tsi

    zeR:

    rtt

    t RTO

    : TCP

    retr

    ansm

    it t

    imeo

    utp:

    stea

    dy-s

    tate

    loss

    eve

    ntra

    te (t

    he d

    iffi

    cult

    par

    t!)

    •w

    ellk

    now

    n ex

    ampl

    e: T

    FRC

    -TC

    P-fr

    iend

    lyra

    teco

    ntro

    l pro

    toco

    l–

    smoo

    th s

    endi

    ngra

    te

    •Ex

    tens

    ion:

    TFM

    CC -

    TCP-

    frie

    ndly

    mul

    tica

    st c

    onge

    stio

    n co

    ntro

    l

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    3636

    Not

    Not

    -- soso-- T

    CPTC

    P-- f

    rien

    dly

    solu

    tion

    sfr

    iend

    ly s

    olut

    ions

    •O

    verc

    ome

    rate

    flu

    ctua

    tion

    s:lim

    it e

    ncod

    ings

    (e.

    g. 2

    or

    3 qu

    alit

    ies)

    , le

    t us

    er d

    ecid

    e

    •Cr

    oss-

    med

    ia-a

    dapt

    atio

    n:ch

    oose

    fro

    m v

    ideo

    , au

    dio,

    sin

    gle

    pict

    ures

    , te

    xt(e

    .g.

    MPE

    G7)

    •Li

    mit

    by

    bott

    lene

    ck b

    andw

    idth

    –of

    ten:

    "la

    st m

    ile"

    -e.

    g. R

    ealM

    edia

    –be

    tter

    : de

    term

    ine

    actu

    al b

    ottl

    enec

    k vi

    a pa

    cket

    pai

    r ap

    proa

    ch

    •If

    wir

    eles

    s lin

    kin

    volv

    ed:

    smal

    l pac

    kets

    , U

    DP

    Lite

    •An

    othe

    r po

    ssib

    ility

    :se

    nd m

    ore

    (do

    FEC)

    in r

    espo

    nse

    to p

    acke

    t lo

    ss–

    (ver

    y ne

    twor

    k-un

    frie

    ndly

    beh

    avio

    ur,

    but

    may

    yie

    ld le

    ss d

    ata

    loss

    )

    Network stability: some

    adaptation better than none!

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    3737

    Som

    e th

    ough

    tsSo

    me

    thou

    ghts

    •H

    ow T

    CP-f

    rien

    dly

    are

    8 w

    eb b

    row

    sers

    ?–

    Cong

    esti

    on M

    anag

    er:

    cong

    esti

    on c

    ontr

    ol f

    or a

    ll fl

    ows

    in O

    S co

    re–

    Mul

    TCP:

    Em

    ulat

    e m

    ulti

    ple

    TCP‘

    s to

    pro

    vide

    dif

    fere

    ntia

    ted

    serv

    ices

    •H

    ow T

    CP-f

    rien

    dly

    are

    shor

    t-liv

    ed f

    low

    s? (

    web

    -tra

    ffic

    , ..

    )

    •H

    ow t

    o co

    nvin

    ce In

    tern

    et m

    ulti

    med

    ia a

    pp.

    prog

    ram

    mer

    s to

    im

    plem

    ent

    TCP-

    frie

    ndly

    con

    gest

    ion

    cont

    rol?

    •So

    luti

    on:

    Dat

    agra

    m C

    onge

    stio

    n Co

    ntro

    l Pro

    toco

    l (D

    CCP)

    –W

    ell-

    defi

    ned

    fram

    ewor

    k fo

    r (T

    CP-f

    rien

    dly)

    con

    gest

    ion

    cont

    rol

    –Se

    nder

    app

    cho

    oses

    an

    appr

    opri

    ate

    cong

    esti

    on c

    ontr

    ol m

    echa

    nism

    –Co

    re O

    S im

    plem

    enta

    tion

    of

    mec

    hani

    sms

    –Lo

    ts o

    f ad

    diti

    onal

    fea

    ture

    s: n

    once

    s, p

    arti

    al /

    sep

    arat

    e ch

    ecks

    ums

    (dis

    ting

    uish

    : co

    rrup

    tion

    ⇔co

    nges

    tion

    ), .

    ..

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    3838

    Het

    erog

    eneo

    us e

    nvir

    onm

    ents

    Het

    erog

    eneo

    us e

    nvir

    onm

    ents

    TCP

    over

    hig

    h sp

    eed

    links

    :

    •la

    rger

    init

    ial w

    indo

    w /

    win

    dow

    sca

    ling

    opti

    on,

    TCP

    SACK

    •Sc

    alab

    le T

    CP:

    incr

    ease

    /dec

    reas

    e fu

    ncti

    ons

    chan

    ged

    (pro

    bing

    tim

    es p

    ropo

    rtio

    nal t

    o rt

    t bu

    t no

    t ra

    te)

    •H

    ighS

    peed

    TCP

    (mer

    ged

    wit

    h Sc

    alab

    le T

    CP):

    resp

    onse

    fun

    ctio

    n le

    ss d

    rast

    ic in

    hig

    h ba

    ndw

    idth

    env

    iron

    men

    ts o

    nly

    •Q

    uick

    -Sta

    rt:

    quer

    y ro

    uter

    s fo

    r in

    itia

    l sen

    ding

    rat

    e w

    ith

    IP o

    ptio

    nsdr

    aft

    only

    -se

    ems

    to b

    e ab

    ando

    ned?

    !

    TCP

    over

    asy

    mm

    etri

    c lin

    ks:

    •AC

    K su

    ppre

    ssio

    n, A

    CK c

    ompa

    ctio

    n, T

    CP h

    eade

    r co

    mpr

    essi

    on

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    3939

    TCP

    over

    noi

    sy (

    wir

    eles

    s) li

    nks

    TCP

    over

    noi

    sy (

    wir

    eles

    s) li

    nks

    •Va

    riou

    s po

    ssib

    le e

    nhan

    cem

    ents

    :–

    split

    con

    nect

    ion

    at b

    ase

    stat

    ion

    –m

    onit

    or c

    onne

    ctio

    n at

    bas

    e st

    atio

    n, b

    uffe

    r +

    inte

    rfer

    e (“

    Snoo

    p TC

    P”)

    •N

    ote:

    ECN

    is n

    ot a

    ffec

    ted

    by li

    nk n

    oise

    !

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    4040

    Beyo

    nd E

    CNBe

    yond

    ECN

    •AT

    M:

    Expl

    icit

    Rat

    e Fe

    edba

    ck(p

    art

    of A

    vaila

    ble

    Bit

    Rate

    (AB

    R)se

    rvic

    e)RM

    (re

    sour

    ce m

    anag

    emen

    t) c

    ells

    :–

    sent

    by

    send

    er,

    inte

    rspe

    rsed

    wit

    h da

    ta c

    ells

    ; bi

    ts in

    RM

    cel

    l set

    by

    swit

    ches

    NI b

    it:

    no in

    crea

    se in

    rat

    e (m

    ild c

    onge

    stio

    n),

    (EF)

    CI b

    it:

    like

    Inte

    rnet

    ECN

    •tw

    o-by

    te E

    R (e

    xplic

    it r

    ate)

    fie

    ld:

    may

    be

    low

    ered

    by

    cong

    este

    d sw

    itch

    •se

    nder

    ’ se

    nd r

    ate

    thus

    min

    imum

    sup

    port

    able

    rat

    e on

    pat

    h!

    •Pr

    oble

    m:

    ATM

    fai

    led

    (sca

    labi

    lity?

    too

    muc

    h co

    mpl

    exit

    y in

    sw

    itch

    es?)

    •Ex

    peri

    men

    tal I

    nter

    net

    appr

    oach

    es:

    •M

    ulti

    leve

    l ECN

    (tw

    o bi

    ts),

    eXp

    ress

    Con

    trol

    Pro

    toco

    l (XC

    P),

    CAD

    PC/P

    TP(m

    y ow

    n)

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    4141

    Oth

    er T

    CP e

    nhan

    cem

    ents

    Oth

    er T

    CP e

    nhan

    cem

    ents

    •FA

    ST T

    CP–

    Vari

    ant

    base

    d on

    win

    dow

    and

    del

    ay–

    Del

    ay a

    llow

    s fo

    r ea

    rlie

    r ad

    apta

    tion

    (aw

    aren

    ess

    of g

    row

    ing

    queu

    e)–

    Prov

    en t

    o be

    sta

    ble

    –Co

    mm

    erci

    ally

    ann

    ounc

    ed +

    pat

    ent

    prot

    ecte

    d, b

    y St

    even

    Low

    ‘s C

    alTe

    ch g

    roup

    –an

    othe

    r de

    lay-

    base

    d ex

    ampl

    e: T

    CP V

    egas

    •TC

    P W

    estw

    ood

    –di

    ffer

    ent

    resp

    onse

    fun

    ctio

    n (p

    ropo

    rtio

    nal t

    o ra

    te)

    –pr

    oven

    to

    be s

    tabl

    e

    •Lo

    ts o

    f ex

    peri

    men

    tal A

    ctiv

    e Q

    ueue

    Man

    agem

    ent

    sche

    mes

    out

    the

    re–

    Adap

    tive

    RED

    , BL

    UE,

    REM

    , RI

    O e

    tc.

    etc.

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    4242

    How

    to

    desi

    gn y

    our

    own

    mec

    hani

    smH

    ow t

    o de

    sign

    you

    r ow

    n m

    echa

    nism

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    4343

    End

    syst

    ems:

    Mea

    sure

    ...

    End

    syst

    ems:

    Mea

    sure

    ...

    ++w

    ell s

    tudi

    ed

    -- lea

    ds t

    o m

    isin

    terp

    reta

    tion

    of t

    rans

    mis

    sion

    err

    ors

    thro

    ughp

    ut("

    good

    put"

    )(

    mea

    n, f

    luct

    uati

    ons,

    pack

    et lo

    ss r

    atio

    ..)

    ++ea

    sy t

    o m

    easu

    re

    ++in

    depe

    nden

    t of

    tra

    nsm

    issi

    on e

    rror

    s

    -- not

    pra

    ctic

    al w

    itho

    ut t

    hrou

    ghpu

    t

    dela

    y(

    rtt,

    one

    way

    del

    ay,

    jitt

    er..

    )

    sim

    ilar

    dela

    y, d

    iffe

    rent

    av

    aila

    ble

    band

    wid

    th!

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    4444

    ...

    and

    chan

    ge!

    ...

    and

    chan

    ge!

    •lo

    wer

    laye

    rs–

    thro

    ughp

    ut(g

    ap b

    etw

    een

    pack

    ets)

    •w

    indo

    w b

    ased

    / r

    ate

    base

    d•

    wel

    l stu

    died

    , m

    any

    opti

    ons

    -ou

    r m

    ain

    inte

    rest

    !–

    pack

    et s

    ize

    •la

    rge:

    rec

    omm

    ende

    dle

    ss o

    verh

    ead

    •sm

    all:

    less

    impa

    ct o

    f tr

    ansm

    issi

    on e

    rror

    s,sm

    alle

    r la

    tenc

    y!–

    prot

    ocol

    Not

    e:

    Not

    e: p

    acke

    t si

    ze =

    gra

    nula

    rity

    of

    thro

    ughp

    ut m

    easu

    rem

    ents

    •co

    nten

    t–

    com

    pres

    sion

    –hi

    erar

    chic

    al e

    ncod

    ing

    –FE

    C

    Com

    mon

    dif

    ficu

    ltie

    s:Co

    mm

    on d

    iffi

    cult

    ies:

    band

    wid

    th k

    now

    n (d

    epen

    ding

    on

    cont

    ent)

    ?gr

    anul

    arit

    y of

    rat

    e ch

    ange

    s

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    4545

    Mea

    suri

    ng t

    he n

    etw

    ork

    Mea

    suri

    ng t

    he n

    etw

    ork

    •W

    hen

    you

    mea

    sure

    , yo

    u m

    easu

    re t

    he p

    ast

    –pr

    edic

    tion

    s /

    esti

    mat

    ions

    wit

    h a

    ?? %

    cha

    nce

    of s

    ucce

    ss o

    r co

    ntro

    l the

    ory

    •W

    hen

    you

    mea

    sure

    , yo

    u ch

    ange

    the

    sys

    tem

    –th

    ink

    of u

    nres

    pons

    ive

    flow

    s vs

    . TC

    P–

    non-

    intr

    usiv

    enes

    s re

    ally

    impo

    rtan

    t (e

    .g.,

    mon

    itor

    TCP

    beh

    avio

    r)

    •M

    easu

    rem

    ents

    yie

    ld n

    o gu

    aran

    tees

    –In

    tern

    et t

    raff

    ic =

    res

    ult

    of u

    ser

    beha

    vior

    !

    •Re

    sear

    ch c

    arri

    ed o

    ut in

    con

    trol

    labl

    e, is

    olat

    ed e

    nvir

    onm

    ents

    –Fi

    eld

    tria

    ls a

    re a

    nec

    essa

    ry e

    xtra

    whe

    n yo

    u kn

    ow t

    hat

    som

    ethi

    ng w

    orks

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    4646

    Poss

    ibili

    ties

    in r

    oute

    rsPo

    ssib

    iliti

    es in

    rou

    ters

    •Co

    mm

    unic

    ate

    wit

    h en

    d sy

    stem

    s–

    alte

    r he

    ader

    fla

    gs (

    IP o

    nly!

    sho

    uld

    not

    look

    at

    othe

    r la

    yers

    )–

    gene

    rate

    sig

    nalin

    g pa

    cket

    s: In

    tern

    et C

    ontr

    ol M

    essa

    ge P

    roto

    col (

    ICM

    P)(m

    ainl

    y er

    ror

    mes

    sage

    s)

    •Co

    ntro

    l pac

    kets

    in q

    ueue

    s:us

    e qu

    eue

    leng

    th o

    r po

    siti

    on in

    que

    ue t

    o–

    com

    mun

    icat

    e („

    mar

    k“)

    –dr

    op,

    mov

    e to

    oth

    er q

    ueue

    etc

    .

    •Pr

    oble

    ms

    –CP

    U c

    ycle

    s sc

    arce

    in (

    core

    ) In

    tern

    et r

    oute

    rs–

    Scal

    abili

    ty!

    (e.g

    ., n

    o pe

    r-fl

    ow s

    tate

    )ex

    ampl

    e: IC

    MP

    Sour

    ce Q

    uenc

    h fa

    iled

    (con

    gest

    ion

    noti

    fica

    tion

    in t

    imes

    of

    cong

    esti

    on)

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    4747

    My

    Ph.D

    . re

    cipe

    :

    My

    Ph.D

    . re

    cipe

    : --

    ))

    Und

    erly

    ing

    thou

    ght:

    “TCP

    alw

    ays

    exce

    eds

    the

    avai

    labl

    e ba

    ndw

    idth

    in

    orde

    r to

    det

    ect

    it

    (whe

    n it

    is

    alre

    ady

    too

    late

    ). W

    ould

    n‘t

    it b

    e be

    tter

    to

    ask

    for

    the

    avai

    labl

    e ba

    ndw

    idth

    ?“

    Proc

    ess:

    •de

    sign

    suc

    h a

    mea

    ns:

    “Per

    form

    ance

    Tra

    nspa

    renc

    y Pr

    otoc

    ol (

    PTP)

    “–

    note

    :m

    ust

    lead

    to

    abso

    lute

    ly g

    reat

    res

    ults

    in o

    rder

    to

    just

    ify

    rout

    er e

    ffor

    t!

    •fi

    nd o

    ut h

    ow t

    o us

    e th

    e (a

    vaila

    ble

    band

    wid

    th)

    info

    rmat

    ion

    ...w

    itho

    ut b

    eing

    a c

    ontr

    ol t

    heor

    y gu

    ru!

    ⇒th

    e tr

    icky

    par

    t!

    •m

    ixtu

    re o

    f in

    tuit

    ion,

    mat

    hs,

    sim

    ulat

    ion,

    ..

    Let‘s look

    at this!

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    4848

    Exte

    nded

    Use

    of

    Vect

    or D

    iagr

    ams

    Exte

    nded

    Use

    of

    Vect

    or D

    iagr

    ams

    •Pr

    oble

    m:

    –St

    abili

    ty a

    naly

    sis

    com

    plex

    –TC

    P-lik

    e m

    echa

    nism

    des

    ign

    diff

    icul

    t

    •So

    luti

    on:

    –Ex

    tend

    ed u

    se o

    f ve

    ctor

    dia

    gram

    s!

    •An

    alyz

    e ac

    tual

    res

    ults

    (fr

    om s

    imul

    atio

    n or

    rea

    l lif

    e m

    easu

    rem

    ents

    )

    •In

    stea

    d of

    jus

    t ex

    plai

    ning

    a c

    once

    pt,

    desi

    gnin

    the

    2D

    dia

    gram

    spa

    ce!

    –N

    eces

    sary

    sim

    plif

    icat

    ions

    may

    eve

    n be

    les

    sdr

    amat

    ic!

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    4949

    How

    Sta

    ble

    is A

    IMD

    / a

    sync

    . RT

    T?H

    ow S

    tabl

    e is

    AIM

    D /

    asy

    nc.

    RTT?

    U 2

    SER

    Use

    r 1

    -0.0

    500

    0.00

    00

    0.05

    00

    0.10

    00

    0.15

    00

    0.20

    00

    0.25

    00

    0.30

    00

    0.35

    00

    0.40

    00

    0.45

    00

    0.50

    00

    0.55

    00

    0.60

    00

    0.65

    00

    0.70

    00

    0.75

    00

    0.80

    00

    0.85

    00

    0.90

    00

    0.95

    00

    1.00

    00

    0.00

    000.

    2000

    0.40

    000.

    6000

    0.80

    001.

    0000

    Fluid model

    (no queues, ..)

    RTT: 7 vs. 2

    AI=0.1, MD=0.5

    Sim. time=175

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    5050

    Is A

    IMD

    dis

    tort

    ed in

    TCP

    ?Is

    AIM

    D d

    isto

    rted

    in T

    CP?

    TCP 2

    TC

    P 1

    1.00

    00

    1.50

    00

    2.00

    00

    2.50

    00

    3.00

    00

    3.50

    00

    4.00

    00

    4.50

    00

    5.00

    00

    5.50

    00

    6.00

    00

    6.50

    00

    7.00

    00

    7.50

    00

    8.00

    00

    8.50

    00

    9.00

    00

    9.50

    00

    10.0

    000

    2.00

    004.

    0000

    6.00

    008.

    0000

    10.0

    000

    12.0

    000

    14.0

    000

    ns-2 simulator

    TCP Tahoe

    equal RTTs

    1 bottleneck link

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    5151

    Vari

    ous

    othe

    r Po

    ssib

    iliti

    esVa

    riou

    s ot

    her

    Poss

    ibili

    ties

    •An

    alyz

    e re

    al li

    fe d

    ata

    •An

    alyz

    e di

    ffer

    ent

    mec

    hani

    sms

    –m

    ore

    com

    plex

    fee

    dbac

    k: A

    TM A

    BR–

    queu

    eing

    beh

    avio

    ur:

    AQM

    –..

    .

    •Pe

    rfor

    m a

    naly

    sis

    in v

    ecto

    r di

    agra

    m s

    pace

    –pl

    ot “

    dist

    ance

    fro

    m o

    ptim

    alit

    y“ /

    tim

    e de

    velo

    pmen

    t–

    plot

    tim

    e of

    con

    verg

    ence

    / u

    ser

    1 al

    loca

    tion

    –..

    .

    •..

    . ve

    ctor

    dia

    gram

    aid

    ed d

    esig

    n!

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    5252

    Inte

    ract

    ive

    Vect

    or D

    iagr

    am A

    ided

    Des

    ign

    Inte

    ract

    ive

    Vect

    or D

    iagr

    am A

    ided

    Des

    ign

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    5353

    A f

    oolp

    roof

    Ph.

    D.

    thes

    is r

    ecip

    eA

    foo

    lpro

    of P

    h.D

    . th

    esis

    rec

    ipe

    •Ro

    ughl

    y fo

    llow

    thi

    s ta

    ble

    from

    1 t

    o 7

    ...

    if s

    omet

    hing

    fai

    ls,

    goba

    ck!

    •In

    my

    case

    :–

    Leve

    l 2:

    intu

    itio

    n (f

    oolin

    g ar

    ound

    wit

    h C

    AVT)

    –Le

    vel 4

    :th

    e “H

    eure

    ka“

    expe

    rien

    ce -

    intu

    itio

    n w

    as r

    ight

    !–

    Leve

    l 1:

    shee

    t of

    pap

    er–

    Leve

    l 3:

    MS

    Exce

    l + C

    AVT

    refi

    nem

    ent

    (foo

    ling

    arou

    nd,

    part

    2)

    –Le

    vel 5

    :th

    e ty

    pica

    l ns

    + du

    mbb

    ell e

    xper

    ienc

    e–

    Leve

    l 6:

    addi

    tion

    al s

    imul

    atio

    ns–

    ...

    that

    was

    goo

    d en

    ough

    . H

    opin

    g to

    rea

    ch le

    vel 7

    soo

    n.

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    5454

    Even

    tual

    ly:

    CAD

    PCEv

    entu

    ally

    : CA

    DPC

    vs.

    TCP

    vs.

    TCP

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    5555

    CAD

    PC v

    s. 3

    TCP

    (+EC

    N)

    CAD

    PC v

    s. 3

    TCP

    (+EC

    N)

    flav

    ors

    flav

    ors

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    5656

    Refe

    renc

    esRe

    fere

    nces

    1. C

    onge

    stio

    n Co

    ntro

    l: a

    qui

    ck in

    trod

    ucti

    on

    •Ra

    j Ja

    in a

    nd K

    . K.

    Ram

    akri

    shna

    n, "

    Cong

    esti

    on A

    void

    ance

    in C

    ompu

    ter

    Net

    wor

    ks w

    ith

    a Co

    nnec

    tion

    less

    Net

    wor

    k La

    yer:

    Con

    cept

    s, G

    oals

    and

    Met

    hodo

    logy

    '',Pr

    ocee

    ding

    s of

    Com

    pute

    r N

    etw

    orki

    ng S

    ympo

    sium

    , W

    ashi

    ngto

    n, D

    . C.

    , Ap

    ril 1

    1-13

    198

    8, p

    p. 1

    34-1

    43.

    •Va

    n Ja

    cobs

    on,

    "Con

    gest

    ion

    Avoi

    danc

    e an

    d Co

    ntro

    l'',

    Proc

    eedi

    ngs

    of A

    CM S

    IGCO

    MM

    198

    8, p

    p. 3

    14-

    329.

    •D

    . Ch

    iu a

    nd R

    . Ja

    in,

    "Ana

    lysi

    s of

    the

    Incr

    ease

    /Dec

    reas

    e Al

    gori

    thm

    s fo

    r Co

    nges

    tion

    Avo

    idan

    ce in

    Co

    mpu

    ter

    Net

    wor

    ks'',

    Jou

    rnal

    of

    Com

    pute

    r N

    etw

    orks

    and

    ISD

    N,

    Vol.

    17,

    No.

    1,

    June

    198

    9, p

    p. 1

    -14.

    •Sa

    lly F

    loyd

    and

    Van

    Jac

    obso

    n, "

    On

    Traf

    fic

    Phas

    e Ef

    fect

    s in

    Pac

    ket-

    Swit

    ched

    Gat

    eway

    s'',

    Inte

    rnet

    wor

    king

    : Re

    sear

    ch a

    nd E

    xper

    ienc

    e, V

    .3 N

    .3,

    Sept

    embe

    r 19

    92,

    p.11

    5-15

    6. E

    arlie

    r ve

    rsio

    n:

    Com

    pute

    r Co

    mm

    unic

    atio

    n Re

    view

    , V.

    21 N

    .2,

    Apri

    l 199

    1.

    •Sa

    lly F

    loyd

    and

    Van

    Jac

    obso

    n, "

    Rand

    om E

    arly

    Det

    ecti

    on G

    atew

    ays

    for

    Cong

    esti

    on A

    void

    ance

    '',

    IEEE

    /ACM

    Tra

    nsac

    tion

    s on

    Net

    wor

    king

    , Au

    gust

    199

    3.

    •K.

    Ram

    akri

    shna

    n, S

    . Fl

    oyd,

    and

    D.

    Blac

    k, "

    The

    Addi

    tion

    of

    Expl

    icit

    Con

    gest

    ion

    Not

    ific

    atio

    n (E

    CN)

    to

    IP'',

    RFC

    316

    8, S

    epte

    mbe

    r 20

    01.

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    5757

    Refe

    renc

    es /

    2Re

    fere

    nces

    /2

    2. P

    robl

    ems

    •Sc

    ott

    Shen

    ker,

    "Fu

    ndam

    enta

    l D

    esig

    n Is

    sues

    for

    the

    Fut

    ure

    Inte

    rnet

    '', IE

    EE J

    ourn

    al o

    n Se

    lect

    ed A

    reas

    in

    Com

    mun

    icat

    ions

    , 13

    , pp

    . 11

    41-1

    149,

    199

    5.

    •Fr

    ank

    Kelly

    , "C

    harg

    ing

    and

    rate

    con

    trol

    for

    ela

    stic

    tra

    ffic

    '', E

    urop

    ean

    Tran

    sact

    ions

    on

    Tele

    com

    mun

    icat

    ions

    , 8.

    pp.

    33-

    37.

    An u

    pdat

    ed v

    ersi

    on is

    ava

    ilabl

    e at

    ht

    tp:/

    /ww

    w.s

    tats

    lab.

    cam

    .ac.

    uk/f

    rank

    /ela

    stic

    .htm

    l

    •Ra

    mes

    h Jo

    hari

    , "M

    athe

    mat

    ical

    Mod

    elin

    g an

    d Co

    ntro

    l of

    Inte

    rnet

    Con

    gest

    ion"

    , SI

    AM N

    ews,

    Vol

    . 33

    , N

    o. 2

    .

    •L.

    Mas

    soul

    ié a

    nd J

    . Ro

    bert

    s, "

    Band

    wid

    th s

    hari

    ng:

    obje

    ctiv

    es a

    nd a

    lgor

    ithm

    s'',

    Proc

    eedi

    ngs

    of IE

    EE

    Info

    com

    199

    9, N

    ew Y

    ork

    City

    , N

    ew Y

    ork,

    21.

    -25.

    3.

    1999

    .

    •Ra

    mes

    h Jo

    hari

    and

    Dav

    id T

    an,

    "End

    -to-

    End

    Cong

    esti

    on C

    ontr

    ol f

    or t

    he In

    tern

    et:

    Del

    ays

    and

    Stab

    ility

    '', IE

    EE/A

    CM T

    rans

    acti

    ons

    on N

    etw

    orki

    ng 9

    (20

    01)

    818-

    832.

    •As

    hraf

    Mat

    raw

    y an

    d Io

    anni

    s La

    bada

    ris,

    "A

    Surv

    ey o

    f Co

    nges

    iton

    Con

    trol

    Sch

    emes

    for

    Mul

    tica

    st V

    ideo

    Ap

    plic

    atio

    ns",

    IEEE

    Com

    mun

    icat

    ions

    Sur

    vey

    & T

    utor

    ials

    , Fo

    urth

    Qua

    rter

    200

    3, V

    ol.

    5, N

    o. 2

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    5858

    Refe

    renc

    es /

    3Re

    fere

    nces

    /3

    3. S

    ome

    prop

    osed

    enh

    ance

    men

    ts

    •M

    ark

    E. C

    rove

    lla a

    nd A

    zer

    Best

    avro

    s, ''

    Self

    -sim

    ilari

    ty in

    Wor

    ld W

    ide

    Web

    Tra

    ffic

    : Ev

    iden

    ce a

    nd

    Poss

    ible

    Cau

    ses''

    , IE

    EE/A

    CM T

    rans

    acti

    ons

    on N

    etw

    orki

    ng,

    Vol.

    5,

    No.

    6,

    Dec

    embe

    r 19

    97.

    •An

    dras

    Ver

    es,

    Zsol

    t Ke

    nesi

    , Sa

    ndor

    Mol

    nar,

    Gab

    or V

    atta

    y, ''

    On

    the

    Prop

    agat

    ion

    of L

    ong-

    rang

    e D

    epen

    denc

    y in

    the

    Inte

    rnet

    '', P

    roce

    edin

    gs o

    f AC

    M S

    IGCO

    MM

    200

    0, S

    tock

    holm

    , Sw

    eden

    , Au

    gust

    28

    -Se

    ptem

    ber

    1 20

    00.

    •G

    uang

    hui H

    e, Y

    uan

    Gao

    , Je

    nnif

    er C

    . H

    ou,

    Kiho

    ng P

    ark,

    ''A

    Case

    for

    Exp

    loit

    ing

    Self

    -Sim

    ilari

    ty o

    f N

    etw

    ork

    Traf

    fic

    in T

    CP'',

    10t

    h IE

    EE In

    tern

    atio

    nal C

    onfe

    renc

    e on

    Net

    wor

    k Pr

    otoc

    ols

    (ICN

    P'02

    ), P

    aris

    , Fr

    ance

    , N

    ov.

    12-1

    5, 2

    002.

    •Jö

    rg W

    idm

    er,

    Robe

    rt D

    enda

    , an

    d M

    arti

    n M

    auve

    , "A

    Sur

    vey

    on T

    CP-F

    rien

    dly

    Cong

    esti

    on C

    ontr

    ol'',

    IE

    EE N

    etw

    ork

    Mag

    azin

    e, S

    peci

    al Is

    sue

    "Con

    trol

    of

    Best

    Eff

    ort

    Traf

    fic'

    ' Vol

    . 15

    , N

    o. 3

    , M

    ay 2

    001.

    •Ph

    ilipp

    e O

    echs

    lin a

    nd J

    on C

    row

    crof

    t, "

    Dif

    fere

    ntia

    ted

    End-

    to-E

    nd In

    tern

    et S

    ervi

    ces

    usin

    g a

    Wei

    ghte

    d Pr

    opor

    tion

    al F

    air

    Shar

    ing

    TCP"

    , AC

    M C

    ompu

    ter

    Com

    mun

    icat

    ion

    Revi

    ew (

    CCR)

    , 19

    98.

    •H

    ari B

    alak

    rish

    nan,

    Har

    ihar

    an R

    ahul

    , an

    d Sr

    iniv

    asan

    Ses

    han,

    "An

    Inte

    grat

    ed C

    onge

    stio

    n M

    anag

    emen

    t Ar

    chit

    ectu

    re f

    or In

    tern

    et H

    osts

    '', P

    roce

    edin

    gs o

    f AC

    M S

    IGCO

    MM

    199

    9, C

    ambr

    idge

    , M

    A.,

    Sept

    embe

    r 19

    99.

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    5959

    Refe

    renc

    es /

    4Re

    fere

    nces

    /4

    •H

    . Ba

    lakr

    ishn

    an a

    nd S

    . Se

    shan

    , "T

    he C

    onge

    stio

    n M

    anag

    er'',

    RFC

    312

    4, J

    une

    2001

    .

    •Ed

    die

    Kohl

    er,

    Mar

    k H

    andl

    ey,

    Sally

    Flo

    yd,

    and

    Jite

    ndra

    Pad

    hye,

    "D

    atag

    ram

    Con

    gest

    ion

    Cont

    rol P

    roto

    col

    (DCC

    P)",

    Inte

    rnet

    -dra

    ft (

    wor

    k in

    pro

    gres

    s) d

    raft

    -iet

    f-dc

    cp-s

    pec-

    05.t

    xt,

    Oct

    ober

    2003

    .

    •M

    etz,

    C.,

    "TC

    P O

    ver

    Sate

    llite

    ...

    The

    Fina

    l Fro

    ntie

    r.",

    IEEE

    Inte

    rnet

    Com

    puti

    ng,

    1999

    .

    •Sa

    lly F

    loyd

    , "H

    ighS

    peed

    TCP

    for

    Lar

    ge C

    onge

    stio

    n W

    indo

    ws"

    ,RF

    C 36

    49,

    Expe

    rim

    enta

    l, D

    ecem

    ber

    2003

    .

    •Ba

    lakr

    ishn

    an,

    H.,

    Pad

    man

    abha

    n, V

    . N

    ., S

    esha

    n, S

    . an

    d Ka

    tz,

    R. H

    ., "

    A Co

    mpa

    riso

    n of

    Mec

    hani

    sms

    for

    Impr

    ovin

    g TC

    P Pe

    rfor

    man

    ce o

    ver

    Wir

    eles

    s Li

    nks"

    , Pr

    ocee

    ding

    s of

    ACM

    SIG

    COM

    M 1

    996,

    Sta

    nfor

    d, C

    A.

    •Ra

    mak

    rish

    nan,

    K.,

    Flo

    yd,

    S. a

    nd B

    lack

    , D

    ., "

    The

    Addi

    tion

    of

    Expl

    icit

    Con

    gest

    ion

    Not

    ific

    atio

    n (E

    CN)

    to

    IP",

    RFC

    316

    8.

    •D

    ina

    Kata

    bi,

    Mar

    k H

    andl

    ey,

    and

    Char

    lie R

    ohrs

    , ''C

    onge

    stio

    n Co

    ntro

    l for

    Hig

    h Ba

    ndw

    idth

    -Del

    ay P

    rodu

    ct

    Net

    wor

    ks'',

    Pro

    ceed

    ings

    of

    ACM

    SIG

    COM

    M 2

    002,

    Pit

    tsbu

    rgh,

    PA,

    19-

    23 A

    ugus

    t 20

    02.

    •Ch

    eng

    Jin,

    Dav

    id X

    . W

    ei a

    nd S

    teve

    n H

    . Lo

    w,

    ''FAS

    T TC

    P: m

    otiv

    atio

    n, a

    rchi

    tect

    ure,

    alg

    orit

    hms,

    pe

    rfor

    man

    ce'',

    IEEE

    Info

    com

    , M

    arch

    200

    4.

  • Uni

    Inn

    sbru

    ckU

    ni I

    nnsb

    ruck

    Info

    rmat

    ik

    Info

    rmat

    ik --

    6060

    Refe

    renc

    es /

    5Re

    fere

    nces

    /5

    4. H

    ow t

    o de

    sign

    you

    r ow

    n m

    echa

    nism

    •M

    icha

    el W

    elzl

    , "V

    ecto

    r Re

    pres

    enta

    tion

    s fo

    r th

    e An

    alys

    is a

    nd D

    esig

    n of

    Dis

    trib

    uted

    Co

    ntro

    ls",

    Pro

    ceed

    ings

    of

    MIC

    200

    2 (I

    ASTE

    D M

    odel

    ling,

    Iden

    tifi

    cati

    on a

    nd C

    ontr

    ol

    Conf

    eren

    ce),

    Inns

    bruc

    k, A

    ustr

    ia,

    18-2

    2 Fe

    brua

    ry 2

    002.

    •M

    icha

    el W

    elzl

    , M

    ax M

    ühlh

    äuse

    r: "

    CAVT

    -A

    Cong

    esti

    on A

    void

    ance

    Vis

    ualiz

    atio

    n To

    ol",

    AC

    M C

    ompu

    ter

    Com

    mun

    icat

    ion

    Revi

    ew,

    Volu

    me

    33,

    Issu

    e 3,

    Jul

    y 20

    03.

    •M

    icha

    el W

    elzl

    : "S

    cala

    ble

    Perf

    orm

    ance

    Sig

    nalli

    ng a

    nd C

    onge

    stio

    n Av

    oida

    nce"

    , Kl

    uwer

    Ac

    adem

    ic P

    ublis

    hers

    , Au

    gust

    200

    3. IS

    BN 1

    -402

    0-75

    70-7

    . Fo

    rew

    ords

    by

    Jon

    Crow

    crof

    t an

    d M

    ax M

    ühlh

    äuse

    r.

    CAVT

    is a

    vaila

    ble

    from

    : ht

    tp:/

    /ww

    w.w

    elzl

    .at/

    tool

    s/ca

    vt/

    Thank you!

    IASTED PDCN 2004 Tutorial:Shaping the Future ofInternet Congestion ControlOutlineCongestion ControlProblem statementA simple router modelThe congestion problemCongestion collapseInternet congestion control: HistoryTCP Congestion Control /1: Tahoe, 1988TCP Congestion Control /2Background: AIMDActive Queue ManagementActive Queue Management /2Explicit Congestion Notification (ECN)ProblemsTCP in heterogeneous environmentsFairnessIssues with TCP-friendlinessProportional FairnessCongestion pricingEnd2end real-time data transferMultimedia adaptationCongestion Control and Quality of ServiceSpecial types of trafficSome reasons for TCP stability“Proofs“ of TCP stabilityUnicast / Broadcast / (overlay) MulticastMu