OSDI 2002 Boston, MA 1 The Effectiveness of Request Redirection on CDN Robustness Limin Wang Vivek...

OSDI 2002 Boston, MA 1

The Effectiveness of Request Redirection on CDN

Robustness

Limin WangVivek Pai and Larry Peterson

Princeton University

December 11, 2002 5th OSDI, Boston, MA 2

Motivation

Server infrastructure criticalFailures are noticeableConsequences are seriousCDNs are commonly used

Robust service under heavy loadNews items, flash crowdsOverload-style attacks (DDoS)

We can exploit CDNs for this


BBB.COM

WWW Service

client


Content Distribution Networks

BBB.COM

client server surrogate

B

B

B

B

B

B

cache

A

A

A

A

A

A

AAA.COM

C

C

C

C

C

C

CCC.COM

redirector


Partial Replication on CDN

BBB.COM


B

B

B

B

A

A

A

A

AAA.COM

C

C

C

B B

A

A

C

C

C

CCC.COM

redirector


Partial Replication on CDN

BBB.COM


B

B

B

B

A

A

A

A

AAA.COM

C

C

C

C

CCC.COM

redirector


Factors Affecting Redirection

GoalsIncrease system throughput under loadReduce response latency perceived by clients

Server loadPick least loaded server

Network proximity Pick closest server

Cache localityPick server just served the object

What’s the tradeoff across different loads?

Often conflict


Contributions

New Class of CDN Redirection Strategies: Fine Dynamic Replication (FDR)

Selectively replicates URLsBalances load, locality, bandwidth and proximity60-91% throughput improvementDeters DDoS, absorbs flash crowds naturally

New Hybrid Simulator: NS-2 + LogSimEnd-to-end — from SYN to sync()Large scale simulations

• 128 servers, 1000+ clients, ~70,000 req/sec


Outline

Motivation and ContributionsDesign Space of Redirection StrategiesEvaluation MethodologyExperimental ResultsConclusions and Ongoing Work


Redirection Decision Dimensions

URL Hashing – decentralized, coherentDifferent hashing schemes

ReplicationMore replicas closer replicasToo many replicas working set overload

Load knowledgeHow accurate?


Background: Hashing Algorithms

Consistent Hashing (CHash)hash(svr1), … hash(svrN)hash(URL)select “closest” server

Highest Random Weight (HRW)hash(URL, svr1), … hash(URL, svrN)sort hash values, get ordered server listselect server with highest hash value


Strategies: Random

Random (Rand)Requests randomly sent to surrogatesBaseline caseAssumes no pathological behavior


Strategies: Static Replication

Replicated Consistent Hashing (R-CHash)Each URL hashed to a fixed # of replicasFor each request, randomly selects one replicaSimilar to Karger’s Consistent Hashing paper

Replicated Highest Random Weight (R-HRW)

Similar to R-CHash, but use HRW hashingMinimizes two URLs having same set of replicas

Approximates best published CDN algorithmsImprove locality over Random


Strategies: Static + Load

Load-Aware Static Replication (fine grained)LR-CHash

Estimates server load at redirectorsPicks up least-loaded server from server set

LR-HRWCounterpart of LR-CHash

Approximates CDN’s best load balancing behavior


Strategies: Dynamic Replication (New)

Coarse Dynamic Replication (CDR)HRW hashing ordered server listWalks list to find acceptably loaded server# of replicas based on server load (estimated locally)

Fine Dynamic Replication (FDR) # of replicas based on URL popularity, tooFDR-Global, reference scheme with global load info

Dynamically balance load and locality


Summary of StrategiesCategory

Strategy Hashing Scheme Dynamic Server set?

Load aware

?CHash HRW

Random Random

StaticR-CHash R-HRW

Static +Load

LR-CHash LR-HRW

Dynamic

CDR FDR

Network Proximity

NPR-CHash NPLR-CHash

NP-FDR


Evaluation Methodology

SimulationsMerge NS-2 and LogSimSYN to sync(), in detail

Work LoadNormal load• 1000 clients replaying Rice Trace

Flash crowds• 10-80% of 1000 clients hitting 10 URLs• Rest clients replaying Rice Trace


Simulation: Topology

PA CA

WA

SD CA

TX

NECO

GA

MAMIIL

R DC

R

R

RR R

R

R

R

R RR

R R

R

R

R

S

C

S

S

S S

S S S

S

CC C

C

C C C

C

C CC

C

– Server, – Client, – RouterS C R

128 servers, 1000 clients


Capacity: Normal Load (64 svrs, 1000 clnts)

9.3

20.418.5

25.4 25.4

32.6 33.2 33.2

0

5

10

15

20

25

30

35

40

Ag

gre

gate

Syste

m C

ap

acity (K

req

s/s

ec) .

Rand R-CHash R-HRW LR-CHash

LR-HRW CDR FDR FDR-Global

StaticStatic

+ Load Dynamic


Latency Measurement Points

9.3

20.418.5

25.4 25.4

32.6 33.2 33.2

0

5

10

15

20

25

30

35

40

Ag

gre

gate

Syste

m C

ap

acity (K

req

s/s

ec) .



StaticStatic

+ Load Dynamic


Latency: Normal Load, (64 svrs, 1000 clnts)

0

20

40

60

80

100

0.1 1 10 100Response Time (Seconds)

Cu

mu

lati

ve P

erc

en

tag

e .

LR-HRW LR-CHash R-HRW R-CHash FDR CDR Rand

Random Max:9.3K req/s


Latency: Normal Load, (64 svrs, 1000 clnts)

0

20

40

60

80

100


Cu

mu

lati

ve P

erc

en

tag

e .

LR-HRW LR-CHash FDR CDR

Static+Load (LR-HRW) Max: 25.4K req/s


Handle Tail Separately (99% file

530KB)

0

20

40

60

80

100


Cu

mu

lati

ve P

erc

en

tag

e .

LR-HRW CDR-Tail-Static CDR

Static+Load (LR-HRW) Max: 25.4K req/s


Server Resource Utilization (64 svrs)

UtilizationScheme

Max Thpt (Kreq/s)

DISK(%) CPU(%)

Mean Stddev

Mean Stddev

Random 9.3 100.0 0.00 21.03 1.36

R-CHash 20.4 99.15 3.89 57.88 18.36

R-HRW 18.5 99.74 1.26 47.88 15.33

LR-CHash 25.4 97.83 12.51 59.48 18.85

LR-HRW 25.4 99.00 5.94 58.43 16.56

CDR 32.6 36.10 25.18 90.07 11.78

FDR 33.2 33.96 20.38 93.86 7.58

FDR-Global

33.2 17.60 15.43 91.93 11.81Under Normal Work Load at Individual Maximum Capacity


Conclusions

New dynamic schemes work best60-91% improvement versus standard CDNWithout sacrificing latencies Scale well, successful attacks more difficult

Bottleneck shifts from Disk to CPU/NetworkDesired – shows selective replication works

More results in the paperNormal workloads, flash crowd/DDoS trafficNetwork Proximity and HeterogeneityLarge file effects


Ongoing Work

PlanetLab deployment (CoDeeN networks)Intel-funded overlay network, academic testing CDNCurrently running on 12 sites using proxy serversEarly stage, many more to do

Traffic monitoring, capacity test, server management …

More simulationsLarger scale, faster simulated serversMore topologies (power-law, etc)More traces, more elaborate client request model


Thank you!

For More Informationhttp://www.cs.princeton.edu/nsg/cdn

AcknowledgementsHP/Compaq/Intel

AlphaServer ES-40, 8GB RAMrx4610 Server, Itanium, 16GB RAM

iMimicDataReactor Proxy Servers


Capacity Scalability

01020304050607080

0 16 32 48 64 80 96 112 128# Servers

Ag

gre

gat

e S

yste

m C

apac

ity

(Kre

qs/

sec)

.

FDR CDR LR-HRW LR-CHashR-HRW R-CHash Rand

Normal Load, 1000 normal clients


Various Flash Crowds (32 servers)

1 victim file, 1KB 10 victim files, avg 6KB

0

5

10

15

20

25

30

35

40

100 200 300 400 500 600 700 800# Intensive Clients (total clients 1000)

Aggre

gate

Syste

m Ca

pacit

y ( K

reqs/s

ec)

.


0

5

10

15

20

25

30

35

40

100 200 300 400 500 600 700 800# Intensive Clients (total clients 1000)

Aggre

gate

Syste

m Ca

pacit

y ( K

reqs/s

ec)

.



Capacity: Flash Crowd, 64 Servers

11.2

19.8 19.8

31 31

37.8 37.8 38.6

0

5

10

15

20

25

30

35

40

45

Ag

gre

gate

Syste

m C

ap

acity (K

req

s/s

ec) .



750 normal clients, 250 intensive clients

Static Static

+ Load Dynamic

OSDI 2002 Boston, MA 1 The Effectiveness of Request Redirection on CDN Robustness Limin Wang Vivek...

Documents

Transcript of OSDI 2002 Boston, MA 1 The Effectiveness of Request Redirection on CDN Robustness Limin Wang Vivek...