Addressing the Challenges of Web Data Transport fileAddressing the Challenges of Web Data Transport...

29
Addressing the Challenges of Web Data Transport Venkata N. Padmanabhan Microsoft Research UW Whistler Retreat December 1998

Transcript of Addressing the Challenges of Web Data Transport fileAddressing the Challenges of Web Data Transport...

Add

ress

ing

the

Cha

lleng

es o

fW

eb D

ata

Tra

nspo

rt

Ven

kata

N. P

adm

anab

han

Mic

roso

ft R

esea

rch

UW

Whi

stle

r R

etre

atD

ecem

ber

1998

Out

line

•C

halle

nges

•So

lutio

ns–

TC

P Se

ssio

n

–Fa

st S

tart

•O

ngoi

ng a

nd F

utur

e W

ork

The

Big

Pic

ture

Inte

rnet

Ser

ver

Sate

llite

Cab

le

Mod

emC

lien

t

Cli

ent

Ser

ver

Goa

l: T

rans

fer

data

fro

m s

erve

rs to

clie

nts

effi

cien

tly

Why

is th

is h

ard?

#1: M

ultip

le in

depe

nden

t com

pone

nts

#2: B

urst

y da

ta tr

ansf

ers

#3: A

cces

s ne

twor

k ch

arac

teri

stic

s

#1: M

ultip

le I

ndep

ende

nt C

ompo

nent

s

Inte

rlea

ved

data

str

eam

⇒ u

ndes

irab

le c

oupl

ing

Imag

e #1

Imag

e #2

stal

l

Imag

e #1

Imag

e #2

Con

curr

ent d

ata

stre

ams

⇒ c

ompe

titio

n

#2: B

urst

y D

ata

Tra

nsfe

rs

Latency

Ban

dwid

th

Tel

net

Bul

k

Web

•D

ownl

oad

time

sens

itive

to la

tenc

y &

ban

dwid

th

•Sh

ared

net

wor

k ⇒

nee

d to

pro

be b

efor

e us

e

Pro

bing

for

ban

dwid

th r

equi

res

time

#1: H

ow to

avo

id c

ompe

titio

n an

dco

uplin

g?•

HT

TP/

1.0

–av

oids

cou

plin

g bu

t not

com

petit

ion

•P-

HT

TP

[PM

94]

–av

oids

com

petit

ion

but n

ot c

oupl

ing

•T

CP

Con

trol

Blo

ck I

nter

depe

nden

ce [

T97

]–

avoi

ds c

oupl

ing

–av

oids

com

petit

ion

at th

e tim

e of

initi

aliz

atio

n bu

tno

t bey

ond

TC

P Se

ssio

n

Dec

oupl

e se

rvic

e m

odel

fro

m tr

ansp

ort a

lgor

ithm

s

Send

erR

ecei

ver

App

licat

ion

App

licat

ion

TC

P Se

ssio

n

TC

PT

CP

TC

PT

CP

TC

PT

CP

Send

er-s

ide

chan

ges

⇒ e

asy

to d

eplo

y in

crem

enta

lly

TC

P Se

ssio

n

TC

P se

ssio

n co

mpo

nent

s

•In

tegr

ated

con

gest

ion

cont

rol

•C

onne

ctio

n sc

hedu

ling

•In

tegr

ated

loss

rec

over

y

Flex

ible

gra

nula

rity

of

inte

grat

ion

(def

ault:

hos

t-pa

ir)

Con

gest

ion

Con

trol

and

Sch

edul

ing

Key

idea

: how

muc

h da

ta, n

ot w

hat d

ata

•U

nifi

ed c

onge

stio

n w

indo

w c

ontr

ols

amou

nt o

fse

ssio

n-w

ide

outs

tand

ing

data

•W

indo

w g

row

th a

nd s

hrin

kage

not

tied

to th

enu

mbe

r of

con

nect

ions

•D

ecou

ple

conn

ectio

n sc

hedu

ling

from

cong

estio

n co

ntro

l

Com

petin

g T

CP

Con

nect

ions

Tim

e (s

econ

ds)

Data transferred (KB)

4 co

ncur

rent

co

nnec

tions

1.5

Mbp

s/50

ms

emul

ated

link

Com

petit

ion

lead

s to

inco

nsis

tent

per

form

ance

0

20406080100

120

140

02

46

Shar

ing

with

TC

P Se

ssio

n

Tim

e (s

econ

ds)

Data transferred (KB)

BSD

/OS

impl

emen

tatio

n

4 co

ncur

rent

conn

ectio

ns

1.5

Mbp

s/50

ms

emul

ated

link

Shar

ing

lead

s to

mor

e co

nsis

tent

per

form

ance

020406080100

120

140

02

46

Inte

grat

ed L

oss

Rec

over

y

21

32

1Se

rver

Clie

nt1

21

2

4

2

Key

idea

: use

pac

ket o

rder

ing

info

rmat

ion

acro

ssco

nnec

tions

to im

prov

e da

ta-d

rive

n lo

ss r

ecov

ery

dupl

icat

e ac

kla

ter

ack

Perf

orm

ance

# cl

ient

s

Download time per client (sec)

Serv

er a

nd c

lient

s co

nnec

ted

via

1.5

Mbp

s/50

ms

link

4 co

ncur

rent

10

KB

tran

sfer

s be

twee

nse

rver

and

eac

h cl

ient

2-3X

red

uctio

n in

dow

nloa

d tim

e

0246810121416

04

812

1620

inde

pend

ent T

CP

conn

ectio

ns

TC

P se

ssio

n

Pack

et L

oss

# cl

ient

s

# packet losses / timeouts

2X r

educ

tion

in p

acke

t los

ses

due

to in

tegr

ated

con

gest

ion

ctrl

.10

X r

educ

tion

in ti

meo

uts

due

to in

tegr

ated

loss

rec

over

y.

04080120

160

200

04

812

1620

inde

pend

ent T

CP

conn

ectio

ns

TC

P se

ssio

n

2X b

ette

r

L L10

X b

ette

r

TT

L: l

oss

T: t

imeo

ut

Sum

mar

y of

TC

P Se

ssio

n

Key

idea

: sep

arat

ion

of T

CP

func

tion

alit

y

Adv

anta

ges

over

inde

pend

ent T

CP

conn

ectio

ns

•Fe

wer

pac

ket l

osse

s

•B

ette

r lo

ss r

ecov

ery

•M

ore

cont

rol o

ver

sche

dulin

g of

dat

a st

ream

s

Adv

anta

ges

over

P-H

TT

P

•N

o co

uplin

g be

twee

n co

ncur

rent

dat

a st

ream

s

•N

ot ti

ed to

spe

cifi

c ap

plic

atio

n

•C

hang

es c

onfi

ned

to s

ende

r si

de

Ban

dwid

th P

robi

ng in

TC

P

•Sl

ow-s

tart

pro

bing

–ex

pone

ntia

l gro

wth

in c

onge

stio

n w

indo

w s

tart

ing

with

a s

ize

of o

ne s

egm

ent

–ac

k cl

ocki

ng a

void

s bu

rstin

ess

•L

inea

r pr

obin

g

•W

hen

is s

low

-sta

rt p

robi

ng in

itiat

ed?

–up

on c

onne

ctio

n st

art u

p

–up

on r

esta

rt a

fter

an

idle

per

iod

•H

ow d

oes

it im

pact

late

ncy?

–n-

segm

ent t

rans

fer

⇒ a

t lea

st lo

g n

RT

Ts

#2: H

ow to

red

uce

cost

of

prob

ing?

•P-

HT

TP

[PM

94]

–av

oid

repe

ated

pro

bing

for

com

pone

nts

of a

sin

gle

Web

pag

e bu

t not

acr

oss

page

s

•4K

slo

w-s

tart

[A

FP98

]

•R

ate-

base

d Pa

cing

[V

H97

]–

smoo

th o

ut u

sing

est

imat

e of

con

nect

ion

rate

–bu

t the

est

imat

e co

uld

itsel

f be

sta

le

TC

P Fa

st S

tart

Bas

ic id

ea:

use

cach

ed n

etw

ork

para

met

ers

tore

duce

the

cost

of

prob

ing

•R

euse

mos

t rec

ent s

ucce

ssfu

l win

dow

siz

e–

slow

-sta

rt ⇒

old

cwnd

/2, l

inea

r ph

ase

⇒ o

ldcw

nd-1

•E

stim

ate

conn

ectio

n’s

rate

as

cwnd

/srt

t

•B

reak

up

larg

e bu

rst i

nto

max

burs

t-si

zed

burs

ts

32

1Se

rver

Clie

nt5

23

1

54

6

46

Dyn

amic

s of

Fas

t Sta

rt

05101520253035

01

23

0

1020304050

01

23

FS TC

P

FS TC

P

Con

gest

ion

win

dow

(K

B)

Sequ

ence

num

ber

(KB

)

Dat

a tr

ansf

er o

ver

Dir

ecPC

sat

ellit

e ne

twor

k

0.95

2.60

Tim

e (s

ec)

Tim

e (s

ec)

Rob

ustn

ess

of F

ast S

tart

Goa

l: F

ast s

tart

sho

uld

help

whe

n ca

ched

info

isva

lid b

ut n

ot h

urt w

hen

it is

sta

le

Stud

ies

indi

cate

that

ava

ilabl

e ba

ndw

idth

is o

ften

stab

le f

or s

ever

al m

inut

es [

P97,

BSS

K97

]

But

we

need

to g

uard

aga

inst

sta

lene

ss

•Pr

otec

ting

othe

rs

•Pr

otec

ting

ones

elf

Prot

ectin

g O

ther

s

•E

nabl

es c

ontr

ol o

n tim

e sc

ale

fine

r th

an R

TT

•A

void

s po

tent

ial c

onge

stio

n co

llaps

e

R

F

S

F

in-p

rofi

le

out-

of-p

rofi

le

Pro

tect

oth

ers

from

ove

r-ag

gres

sive

fast

sta

rt

Pref

eren

tially

dro

p fa

st s

tart

pac

kets

(ex

cept

fir

st o

ne)

Prot

ectin

g O

nese

lf

Pro

tect

one

self

from

con

sequ

ence

s of

bur

st lo

ss

Qui

ckly

det

ect a

nd a

bort

fai

led

fast

sta

rt a

ttem

pt

•Fi

ne-g

rain

ed r

eset

tim

er d

urin

g fa

st s

tart

pha

se–

tied

to th

e fa

st T

CP

tim

er (

200

ms)

•If

res

et ti

mer

exp

ires

, abo

rt f

ast s

tart

–re

set c

wnd

to o

ne s

egm

ent,

initi

ate

slow

sta

rt

–no

oth

er c

onge

stio

n co

ntro

l pen

altie

s•

ssth

resh

not

hal

ved,

RT

O n

ot b

acke

d of

f

•A

bort

als

o w

hen

mul

tiple

loss

es w

ithin

RT

T

Impa

ct o

f St

alen

ess

on O

nese

lf

Abo

rtin

g fa

st s

tart

in c

ase

of f

ailu

re p

reve

nts

sign

ific

ant p

erfo

rman

ce d

egra

datio

n1.5

Mbp

s50

ms

051015202530

02

46

810

# o

f co

mp

etin

g b

ulk

tra

nsf

ers

Download Time (seconds)

TC

PFS FS

-CO

AR

SEFS

-NO

-AB

OR

T

Impa

ct o

f St

alen

ess

on O

ther

s

Prio

rity

dro

ppin

g si

gnif

ican

tly d

ecre

ases

adv

erse

impa

ct o

n co

mpe

ting

traf

fic

2150

2200

2250

2300

2350

2400

2450

2500

2550

1920

2122

2324

25

Tim

e (s

ec)

Segment #

FS FS w

/o P

-Dro

p

Asy

mm

etri

c A

cces

s N

etw

ork

•Pr

oble

m: u

pstr

eam

dat

a pa

cket

s bl

ock

acks

–R

TT

can

bec

ome

very

larg

e

•Po

ssib

le s

olut

ion:

ack

s-fi

rst s

ched

ulin

g [B

PK97

]–

but R

TT

can

stil

l be

larg

e du

e to

the

pack

et in

tran

smis

sionW

WW

W

Serv

erC

lient

WW

EE

EE

E

Impa

ct o

f B

idir

ectio

nal T

raff

ic

Fast

sta

rt h

elps

eve

n th

ough

the

inhe

rent

RT

T is

not l

arge

051015202530

FIF

OA

cks-

firs

t

TC

PF

S

Download Time (seconds)

175

KB

pag

edo

wnl

oad

over

10 M

bps/

28.8

Kbp

sne

twor

k

Sum

mar

y an

d C

oncl

usio

ns

•T

CP

Sess

ion

–de

coup

les

serv

ice

mod

el f

rom

tran

spor

t alg

orith

ms

–en

able

s co

ncur

renc

y w

ithou

t com

petit

ion

•Fa

st S

tart

–ex

ploi

ts d

iffe

rent

iate

d se

rvic

es to

com

plem

ent e

nd-

to-e

nd c

ontr

ol w

ith f

aste

r tim

e-sc

ale

cont

rol

–im

prov

es b

andw

idth

util

izat

ion

in th

e co

mm

on c

ase

–av

oids

ris

k of

per

form

ance

deg

rada

tion

in th

e w

orst

case

Con

clus

ions

•Fa

st s

tart

is r

obus

t–

sign

ific

ant b

enef

it (2

X)

in f

avor

able

con

ditio

ns

–lit

tle p

erfo

rman

ce d

egra

datio

n in

adv

erse

con

ditio

ns•

prio

rity

dro

ppin

g, q

uick

det

ectio

n of

fai

led

fast

sta

rt

•R

educ

ed la

tenc

y he

lps

both

clie

nts

and

serv

ers

–cl

ient

: fas

ter

dow

nloa

ds

–se

rver

: res

ourc

es f

reed

up

mor

e qu

ickl

y

•Si

gnif

ican

t ben

efit

with

new

acc

ess

netw

orks

–sa

telli

te, c

able

mod

em

–pr

ovid

es p

ath

for

incr

emen

tal d

eplo

ymen

t