outline design and analysis of overlay networkseeweb.poly.edu/el933/slides/class11.pdf · design...

11
1 Design and Analysis of Overlay Networks EL 933, Class11 Yong Liu 11/29/2005 2 Outline ! Paper 1: "Application Level Relay for High-bandwidth Data Transport”, Yong Liu, Yu Gu, Honggang Zhang, Weibo Gong and Don Towsley, the First Workshop on Networks for Grid Applications (GridNets) , October 2004 ! Paper 2: "On the Interaction Between Overlay Routing and Traffic Engineering'', Yong Liu, Honggang Zhang, Weibo Gong and Don Towsley, In Proc. of IEEE/INFOCOM 2005 ! Re-Cap of All Lectures 3 Elements in Computer Networks ! Performance metrics " throughput " delay " reliability ! End Hosts, Routers, Links ! Network Control " routing o OSPF, MPLS, BGP " congestion control o TCP A B 4 Application-level Overlay Networks ! routing overlay: RON, Detour ! content distribution: Akamai ! multicast overlay: Overcast ! computing grids ! p2p file sharing ! appl.-level view of network ! appl.-level control " routes " services A D B C underlay overlay G=(V,E) A D B C

Upload: ngotruc

Post on 06-Feb-2018

216 views

Category:

Documents


0 download

TRANSCRIPT

1

Design and Analysis of Overlay

Networks

EL 933, Class11

Yong Liu

11/29/2005

2

Outline

! Paper 1: "Application Level Relay for High-bandwidth

Data Transport”,

Yong Liu, Yu Gu, Honggang Zhang, Weibo Gong and

Don Towsley, the First Workshop on Networks for

Grid Applications (GridNets) , October 2004

! Paper 2: "On the Interaction Between Overlay

Routing and Traffic Engineering'',

Yong Liu, Honggang Zhang, Weibo Gong and Don

Towsley, In Proc. of IEEE/INFOCOM 2005

! Re-Cap of All Lectures

3

Elements in Computer Networks

! Performance metrics" throughput

" delay

" reliability

! End Hosts, Routers, Links

! Network Control" routing

o OSPF, MPLS, BGP

" congestion control

o TCP

A

B

4

Application-level Overlay Networks

! routing overlay: RON, Detour

! content distribution: Akamai

! multicast overlay: Overcast

! computing grids

! p2p file sharing

! appl.-level view of network

! appl.-level control

" routes

" services

A D

B C

underlay

overlay

G=(V,E)

A D

B C

5

Yet Another Layer of Control?

! Network control

" network-wide performance

" trade off between performance and

• complexity, reliability, manageability, policies

" imperfect practice: network wide and individuals

! Application-level control

" do it on my own

" on top of existing network control

" improved performance for individuals

• throughput, delay, resilience, services.

" selfish behavior v.s. common good

6

Application Level Relay for High-

bandwidth Data Transport

Yong Liu, Yu Gu, Honggang Zhang,

Weibo Gong and Don Towsley

7

Application-level TCP Relay

A B

R R

R R

R

8

Related Work! “Implementation and Performance Evaluation of

Indirect TCP”, Bakre et al. IEEE Trans. Comput., 46

(3), 1997

! “Split TCP for Mobile Ad Hoc Networks”, Kopparty et

al, Symposium on Ad-Hoc Wireless Networks, 2002

! “ROMA: Reliable Overlay Multicast with Loosely

Coupled TCP Connections”, Kwon et al, INFOCOM04

! “Scalability of Reliable Group Communication Using

Overlays”, Baccelli et al, INFOCOM04

9

TCP Pipeline along End-end Path

!TCP performs poorly

for long-haul data

transfers

" low bandwidth

" inefficient packet

retransmission

" long feedback delay

" multiple congested links

0 1 2 n3R R R R

TCP1 TCP2 TCP3 TCPn

!TCP Pipeline: chain of

sequential TCP

connections

" higher bandwidth

" local recovery of

lost packets

" shorter feedback delay

" isolate congested links10

TCP Pipeline Setup

! Pipeline throughput

" homogeneous case: times faster than end-end TCP

" general case: throttled by slowest segment

! Relay nodes incur processing & memory overhead

! Problem: How many relays needed and where?

" find optimal pipeline organization with m (<=n) relays

" given link statistics (delay, loss) solve dynamic programming

problem bottom-up

0 1 2 n3

11

Evaluation: TCP Pipeline! Dynamic Programming

" optimal TCP Pipeline with m relay nodes, (0<=m<=5)

Number of Relays

Thro

ughpu

t (k

bps

)

0 1 2 63 4 540 80 120 40 40 40

12

Pipelining + Striping

! Striping: employ parallel TCP connections to broaden pipelinebottle-neck

! Striping overhead:

" connection setup/packet reassembly/competition among conn.s

! Optimally organize TCP connections sequentially and in parallel

" how many relays and how many TCP connections on each hop?

" dynamic programming problem

0 1 2 n3

13

Setting up Relays Deviating Underlay Path

! Relay path: path in overlay

graph; one TCP connection

per overlay link.

! Path width: minimum TCP

throughput on overlay links

! View TCP throughput as overlay

link weight, find widest path in G

" solved by variant of Dijkstra

shortest path algorithm

overlay

A D

B C

G=(V,E)

underlay

14

Richer Optimality Criteria?

! Tie in the width of relay paths

" slowest overlay link dominates path width

! Additional metric—path length

" underlay hop count: underlay resource

" overlay hop count: relay overhead

! Multi-metric routing problem

" Trade off between path width and length

" Shortest Widest Path

" Widest Path with Hop Constraint

15

Shortest Widest Relay Path

! Shortest-widest relay path: shortest

among all widest relay paths

" find one widest path,

T: width of widest path

" prune overlay links where TCP

throughout < T

" find shortest path in pruned graph

C

B

D

E

A F

108

6

4

10

16

Widest Relay Path with Hop Constraint

! Limit number of relay hops:

" find widest relay path from source s

to any other node in G with at most m

hops

! Solve with variant of Bellman-Ford

algorithm

" assume knowledge of TCP throughput

between adjacent relay nodes, find

widest m-hop relay paths in m

iterations

" at iteration k (<=m), find widest k-hop

relay paths from s to all other nodes

in G

C

B

D

E

A F

108

6

4

10

m=0

m=1

m=2

m=3,4

17

Shortest Path Satisfying Width

Requirement

! Balance between bandwidth and

length

" link/path bandwidth varies over

time, difference between two

relay paths may be small

" threshold based searching:

shortest relay path satisfying

width requirement

" width threshold can be calculated

from widest path search, use

threshold for prune-search

procedure

C

B

D

E

A F

10

98

4

10

?

18

Evaluation: widest relay path with hop

constraint! Underlay: GT-ITM topology, 50 nodes, 217 edges,

random link loss between 0 and 0.02

! Overlay: randomly picked 20 overlay nodes

! Variant of Bellman-Ford algorithm

relay hops

rela

y th

rupt

. (kb

ps)

balancedrelay path

19

Evaluation: shortest relay path satisfying

width requirement! Previous topology

! Find shortest-widest path

! Gradually relax width requirement, find shortest

width-constrained path

Path Width (kbps)

Path

Len

gth

1

2

34

5

balanced

relay path

20

Conclusions & Open Issues

! TCP Pipelining improves throughput of long-haul data

transport, more responsive congestion control

! Optimal relay path established by multi-metric

application-level routing, trade off b.w. and resource

! Future directions

" implementation in overlay networks

• collect information about underlay networks

• routing calculation/update/maintenance

" experiments in real network environment (PlanetLab)

" other applications: wireless/sensors networks

21

On the Interaction Between Overlay

Routing and Traffic Engineering

Yong Liu, Honggang Zhang,

Weibo Gong and Don Towsley

22

! Underlay Routing" determine routes for all source-destination pairs

" minimize network wide delay, congestion, etc.

Routing in Underlay Network

pair 1: A->B

pair 2: A->C

pair 3: C->BA

C

E

B

D

23

! Overlay Routing" choose routes at appl. level

" generates demands forunderlay to carry

" individually vs. cooperatively

! Advantages" better path: delay, loss,

thrupt.

" more responsive

" get around inter-domainrouting policies

Routing in Overlay Network

A

C

EB

D

A

C

B

24

! Overlay Routing" choose routes at appl. level

" generates demands forunderlay to carry

" individually vs. cooperatively

! Advantages" better path: delay, loss,

thrupt.

" more responsive

" get around inter-domainrouting policies

Routing in Overlay Network

Net1Net1

Net2Net2

Net3Net3

A

C

B

! Selfish overlay behavior " impact on overall network performance?

" impact on underlay traffic performance?

25

Related Work

! “On Selfish Routing in Internet-like Environments”,

L. Qiu, Y. R. Yang, Y. Zhang, and S. Shenker, ACM/SIGCOMM,

August 2003

! "Can ISPs Take the Heat from Overlay Networks?”,

R. Keralapura, N. Taft, C. N. Chuah, and G. Iannaccone,

ACM/HotNets-III, November 2004

26

Interactions Between

Overlay Routing and Underlay Routing

Overlay Routing Optimization

minimize overlay cost

Underlay Routing Optimization

minimize underlay cost

overlay

routes

overlay

traffic

underlay

routes

non.over.

traffic

iterations

! equilibrium: existence? uniqueness?

! convergence? oscillations?

! performance of overlay and underlay traffic?

27

Routing Optimization at Two Levels

X: overlay routes Y: underlay routes

! Optimal Overlay Routing

" performance of overlay users

" delay, loss, and congestion etc.

" determined by routes at two levels

" routes satisfying overlay demands

" overlay level flow conservation

" physically carried by underlay

! Optimal Underlay Routing

" performance of all users

" delay, loss, and congestion etc.

" determined by routes at two levels

" routes satisfying all demands

" underlay level flow conservation

" constrained by link capacities

28

Simulation Study! 14 node tier-1 POP network,

! 3 overlay nodes,

! bimodal traffic demand between node pairs

! both overlay and underlay minimize delay

! use LP-Solve to solve for overlay and underlay routes

29

Simulation Study

iteration iteration

average delay of all users average delay of overlay users

traffic demand 1, with 8.1% overlay traffic

after computing underlay routesafter computing overlay routes

perc

enta

ge %

perc

enta

ge %underlay performance

degraded

overlay performance

improved

30

Simulation Study

iteration iteration

traffic demand 2, with 10.8% overlay traffic

perc

enta

ge %

perc

enta

ge %

after computing underlay routesafter computing overlay routes

underlay performance

degraded

overlay performance

degraded

average delay of all users average delay of overlay users

31

Underlay Perf. Degrad. as Func. of

Fraction of Overlay Traffic

!overlay triggers largest oscillations when it

takes about half of total traffic

user

del

ay inc

reas

e

percentage of overlay traffic32

Game Theoretic Study

!Two players non-zero sum gameOverlay

v.s.Underlay

! Repeated Nash game

! Nash Equilibrium Point (NEP)

33

Overlay Routing Model: selfish routing

A B

C! Overlay user individually finds

minimum delay path

" in equilibrium, equal user delays onactive paths

! Underlay minimizes delay of allnetwork users

" in optimum, equal link delayderivatives

! Equilibrium impossible?

34

Overlay Routing Model: selfish routing

A B

C

routing game converges to INEFFICIENT equilibrium!!

! Overlay user individuallyfinds minimum delay path" in equilibrium, equal user

delays on active paths

! Underlay minimizes delay ofall network users" in optimum, equal link delay

derivatives

! Equilibrium exists!

unique equilibrium

mappings from overlay routes to underlay routes resolve conflict

35

Overlay Routing Model: coordinated

routing

! One entity calculates routes for all overlay users

! It knows underlay topology and background traffic

! Given current underlay routing, it solves for

overlay routes explicitly

A B

C

X(k)?

x(k): overlay’s flow on path ACB after round k

36

! There exists unique NEP x*,

! NEP globally stable: x(k) !x*, from any initial x(0)

! Overlay performance degrades for some x(0)

Convergence of Coordinated Overlay

Routing

iteration k iteration k

Overlay Routing Evolution Overlay Delay Evolution

x(k)

x*x(k)<x(k+1)<x*

dela

y Underlay’s turn

Overlay’s turn

37

! There exists unique NEP x*,

! NEP globally stable: x(k) !x*, from any initial x(0)

! Overlay performance degrades for some x(0)

Convergence of Coordinated Overlay

Routing

Round k Round k

x(k)

x*

Underlay’s turn

Overlay’s turn

BAD INTERACTIO

N!

x(k)>x(k+1)>x*

x(k)<x(k+1)<x*

Overlay Delay Evolution

dela

y

Overlay Routing Evolution“best”

strategy

38

Will More Information Help Overlay?

! Overlay knows underlay network topology and its routingstrategy;

! Stackelberg game strategy for overlay routing

" evaluates each overlay routing by predicting underlaynetwork response;

" chooses the one to minimize overlay cost

! Advantage

" one shot game, no oscillation

" best performance for overlay

! Bi-level programming

" NP-hard

" gradient projection search with random start

39

Conclusions & Open Issues

! Selfish overlay routing degrades performance of

network as a whole

! Interactions between blind optimizations at two

levels may converge to lose-lose situation

! Future work:

" larger topology: analysis/experimentation

" overlay routing and inter-domain routing

" interactions between multiple overlays

" implications on design overlay routing

" regulation between overlay and underlay

40

Overlay: Good or Bad Patch to Internet?

! overlays are selfish" overlay routing degrades perf. of whole network

• unfairness to regular users?

" interactions between blind optimizations at two levels mayconverge to lose-lose situation

• overlay friendly underlay?

! overlays are adaptive" overlays respond fast to congestion/anomaly

• win-win situation?

" overlays make underlay less fragile• conjecture: if applications are able to adapt their routes

through the network, the underlay can be less careful inchoosing “optimal” routes

41

We Covered …! Traffic Analysis

" traffic statistics: packets arrive according to Poisson?

#failure of Poisson (LAN,WAN,WWW): LRD, Self-similarity

" how user behavior affect traffic?

#users rule: arrival, thinking time, pkt. inter-arrival

! End-end path characterization

" estimate e2e delay, loss, throughput

#loss not i.i.d; delay: heavy-tail; wireless loss: multi-path

! Network Tomography

" from edge-based traffic measurements, infer internal link-level loss,delay and utilization

# dispersion: link capacity/available bw/bottle-neck

# correlation: link loss/delay, topology

42

We Covered …

! Anomaly Detection" DDos attacks/worm spreading/link failures

# network telescope, signal analysis, reverse eng.

! P2P Measurement" P2P topology/user behavior/workload

# p2p content distribution, traffic locality, ISP, users

! Traffic Matrix (TM) Estimation" existent and new approaches

# under-determined, estimation, new information

43

We Covered …! Optimal Routing

" optimize routes for single TM

# traffic engineering, OSPF weights

" optimal routing in a changing world

$ papers available…

! Congestion Control

" TCP dynamic models, Active Queue Management (AQM)

# TCPs solve a distributed optimization problem

" closed loop analysis

# network modeled by coupled differential equations

% Overlay Networks

" good patches to Internet?

# improved perf.; underlay less fragile

" bad patches to Internet?

# selfish users; bad interaction44

What is next?

! Student presentation

" 12/06/2005

" 12/13/2005

" ?

! Continuing Study

" independent study/project

" papers

! Feedbacks

" how you like it? how I did?

" [email protected]