system & network reading group on selfish routing in internet-like evironments lili qiu...
DESCRIPTION
System & Network Reading Group Motivation (Cont.) Theory front –Roughgarden et al. showed selfish routing can result in serious performance degradation due to lack of cooperationTRANSCRIPT
System & Network Reading Group
On Selfish Routing In Internet-Like Evironments
Lili Qiu (Microsoft Research)Yang Richard Yang (Yale University)
Yin Zhang (AT&T Research)Scott Shenker (ICSI)
System & Network Reading Group
Motivation• Practical front
– Recent studies (e.g., Detour/RON) showed that default routing path is often sub-optimal
– Possible causes of routing inefficiency• Routing hierarchy• Routing policy• Different routing objectives used by ISPs• Stability problem in routing protocols, such as BGP …
– A recent trend: end hosts choose routes• Source routing (e.g., Nimrod)• Overlay routing (e.g., Detour or RON)
– Characteristics of routing by end hosts• Improve over today’s IP routing (e.g., delay, loss rate)• Selfish by nature (i.e., optimize user-centric performance
without considering system-wide criteria)
System & Network Reading Group
Motivation (Cont.)• Theory front
– Roughgarden et al. showed selfish routing can result in serious performance degradation due to lack of cooperation
System & Network Reading Group
Example: Selfish Routing May Yield Sub-Optimal Performance
• Selfish routing– All traffic go through the lower link– Total latency = 1
• Optimal routing (i.e., minimize total latency)– Traffic split equally between the two links– Total latency = ¾
• The performance degradation can be unbounded for non-linear latency functions
src dest
L(x)=1
L(x)=x
System & Network Reading Group
Open Issues• How does selfish routing perform in Internet-like
environments?– Realistic network topologies– Realistic traffic demands– Realistic network delay functions
• How does selfish overlay routing perform?• How does selfish traffic co-exist with the
remaining traffic that uses traditional routing protocols?
• How does users’ selfish routing interact with underlying network control process (e.g., traffic engineering)
System & Network Reading Group
Outline• Overview• Network model• Evaluation Methodology• Performance results
– Physical routing– Overlay routing– Multiple overlays– Interaction with traffic engineering
• Summary and future work
System & Network Reading Group
Overview• Approach
– Use a game-theoretic approach to answer the above open issues
– Focus on intra-domain scenarios•Recent advances in topology mapping and
traffic estimation •Compare with theoretical results
– Focus on equilibrium behavior•Compare the performance of traffic equilibria
with the global optima and default IP routing•Based on realistic topologies, traffic
demands, latency functions
System & Network Reading Group
Network Model• Physical network
– Directed graph G=(V,E)– Latency of each edge is a function of its load (e.g., M/M/1)
• Demands– demand(i,j): the amount of traffic from a source i to a
destination j• Overlays
– A set of overlay nodes, overlay links, and a set of demands– The physical route corresponding to an overlay link is
dictated by network-level routing– Consider mesh-like overlay topologies
• Users– Each user decides how its traffic should be routed– Objective: min latency
System & Network Reading Group
Network Model (Cont.)Route controller
– Uses network-level routing• OSPF: shortest-path with equal-weight splitting, with
the following weight settings– Hop-count– Random-weight– Optimized-compliant weight: minimize network
cost when assuming all traffic is compliant (i.e., following the routes determined by the network) [FRT02]
» Network cost: a piece-wise linear convex function of network load over all links
• MPLS: general multi-commodity flow routing
System & Network Reading Group
Evaluation Methodology• Network topology
– A large tier-1 ISP topology, referred as ISPTopo– Rocketfuel topologies– Random power-law topologies
• Traffic demands– Real traffic demands from the ISPTopo– Synthetic traffic demands
• Link latency functions– M/M/1, M/D/1, P/M/1, P/D/1, BPR
• Performance metrics– Average latency– Maximum link utilization– Network costs: piece-wise linear, increasing, convex
function [FRT02]
System & Network Reading Group
Different Routing Schemes• Physical routing
– Source routing (i.e., selfish routing studied in previous theoretical work)
– Optimal routing• Overlay routing
– Overlay source routing (i.e., selfish routing with routing constraints)
– Overlay optimal routing• Compliant routing (i.e., normal Internet
routing)
System & Network Reading Group
Approach to Computing the Traffic Equilibria
• General approach– Simulation-based: too expensive– We use a game-theoretic approach to compute the traffic
equilibria directly• Computing the equilibria of physical routing
– linear-approximation algorithm, a variant of Frank-Wolfe algorithm
• Computing the equilibria of overlay routing– Symmetric: Modified linear approximation algorithm – Asymmetric: Jacob’s relaxation algorithm
• Computing the equilibria of multiple overlays– Use the relaxation algorithm to guarantee the
convergence
System & Network Reading Group
Outline• Overview• Network model• Evaluation Methodology• Performance Evaluation
– Source routing– Overlay routing– Multiple overlays– Interaction with traffic engineering
• Summary and future work
System & Network Reading Group
Selfish Source Routing• Questions
– Are Internet-like environments among the worst-case?
– What is the system-wide cost for selfish source routing?
• Dimensions – Performance metrics: latency & network load– Effects of network topologies– Effects of network load– Effects of latency functions
System & Network Reading Group
Selfish Source Routing: Latency• Effects of network topologies (M/M/1, load
scale factor=1, OC3 bandwidth)
0
5000
10000
15000
20000
25000
Abo
vene
t
ATT
EB
ON
E
Exo
dus
Leve
l3
Spr
int
Tels
tra
Tisc
ali
Ver
io
Pow
erD
2
Pow
erD
5
Pow
erD
10
Load scale factor=1
Ave
rage
late
ncy
(us)
source optimal compliant
Selfish routing yields close to optimal latency, much better than compliant routing
System & Network Reading Group
Selfish Source Routing: Network Load
• Effects of network topologies
1
10
100
1000
10000
Abo
vene
t
ATT
EB
ON
E
Exo
dus
Leve
l3
Spr
int
Tels
tra
Tisc
ali
Ver
io
Pow
erD
2
Pow
erD
5
Pow
erD
10
Load scale factor=1Ne
twor
k co
st
source optimal compliant
020406080
100120140160
Abo
vene
t
ATT
EB
ON
E
Exo
dus
Leve
l3
Spr
int
Tels
tra
Tisc
ali
Ver
io
Pow
erD
2
Pow
erD
5
Pow
erD
10
Load scale factor=1
Max
imum
link
util
izat
ion
(%)
source optimal compliant
Selfish routing tends to overload links.
System & Network Reading Group
Summary: Selfish Source Routing• The performance is qualitatively the same
as we vary latency functions and network load
• Unlike the theoretical worst cases, selfish source routing yields close to optimal latency
• Selfish routing tends to overload links on the shortest paths
System & Network Reading Group
Outline• Overview• Network model• Evaluation Methodology • Performance results
– Source routing– Overlay routing– Multiple overlays– Interaction with traffic engineering
• Conclusion and future work
System & Network Reading Group
Selfish Overlay Routing• Questions
– Does selfish overlay routing perform well?
– How does the coverage of overlay network affect the performance?
• Dimensions– Effects of network topologies– Effects of amount of overlay coverage– Effects of how overlay nodes are
selected (e.g., random or edge nodes)
System & Network Reading Group
Difference between Source Routing and Overlay Routing
• Even if the overlay includes all network nodes, routing on an overlay is still different – Network-level routing can prevent overlay traffic
from using a link by setting the corresponding entry in routing matrix to 0 (in OSPF this is achieved by assigning a large weight)
– Certain physical routes cannot be implemented by any overlay routing
• Routing flexibility is further reduced when only a fraction of nodes belong to an overlay
System & Network Reading Group
Selfish Overlay Routing (Full Overlay Coverage)
0
50
100
150
200
0 0.5 1 1.5 2
Load scale factor
Max
link
util
izat
ion
source routing overlay-src: opt-weight
overlay-src: hop-count overlay-src: rand-weight
02000400060008000
1000012000
0 0.5 1 1.5 2
Load scale factor
Aver
age
late
ncy
(us)
source routing overlay-src: opt-weight
overlay-src: hopt-count overlay-src: rand-weight
1) overlay-src with opt-weight and hop-count performsimilarly as source routing
2) overlay-src with random-weight performs much worse.
System & Network Reading Group
Selfish Overlay Routing (Full Overlay Coverage)
• Direct Link Shortest [DLS]– For any physically adjacent nodes A and B, all the
traffic from A to B is routed through the direct link AB without involving any other links. (e.g., hop-count-based OSPF)
• For an overlay that covers all network nodes and satisfies DLS– routing on the overlay = routing on the underlay
• Hop-count-based OSPF and optimized OSPF weights satisfy DLS they perform similarly as source routing
• Random OSPF weights violate DLS some links are pruned, and performance degrades
System & Network Reading Group
Selfish Overlay Routing (Partial Overlay Coverage)
• Overlay is formed from all edge nodes in ISPTopo
ISPTopo
7500
8000
8500
9000
9500
0 0.5 1 1.5 2 2.5
Load scale factor
Ave
rage
late
ncy
(us)
all partial
ISPTopo
0
50
100
150
200
0 0.5 1 1.5 2 2.5
Load scale factorM
ax li
nk u
tiliz
atio
n (%
)
all partial
The effects of partial overlay coverage is smallin backbone topologies.
System & Network Reading Group
Summary: Selfish Overlay Routing
• For full overlay coverage– Overlay has full routing control when the
underlay satisfies DLS– The only way in which OSPF affects
overlay routing is by violating DLS, which could reduce available network resources
– Overlay source routing reduces latency at the expense of higher network cost
• The effects of partial overlay coverage are small in backbone topologies
System & Network Reading Group
Outline• Overview• Network model• Evaluation Methodology • Performance results
– Source routing– Overlay routing– Multiple overlays– Interaction with traffic engineering
• Conclusion and future work
System & Network Reading Group
Interactions among Competing Overlays
• Question– Can multiple overlays share network
resources fairly and effectively?• Dimensions
– Effects of network topologies– Effects of network-level routing
schemes– Effects of network load and traffic
distribution among overlays– Effects of the number of competing
overlays
System & Network Reading Group
Interactions among Competing Overlays (Cont.)
• Effects of network-level routing load scale factor = 1
05000
100001500020000250003000035000
com
/com
opt/c
omop
t/opt
src/
com
src/
opt
src/
src
com
/com
opt/c
omop
t/opt
src/
com
src/
opt
src/
src
com
/com
opt/c
omop
t/opt
src/
com
src/
opt
src/
src
opt-comp hop-count random
Ave
rage
late
ncy
(us)
foreground background
System & Network Reading Group
Summary: Interactions among Competing Overlays
• With reasonable OSPF weights (e.g., hop-count)– Different routing schemes co-exist
without hurting each other• With bad OSPF weights
– Selfish overlay improves both for themselves and for compliant traffic
System & Network Reading Group
Outline• Overview• Network model• Evaluation Methodology • Performance results
– Source routing– Overlay routing– Multiple overlays– Interactions with traffic engineering
• Conclusion and future work
System & Network Reading Group
Selfish Routing vs. Traffic Engineering
• So far we assume network is dumb (i.e., static underlay routing)
• In practice, the network is smart due to traffic engineering (i.e., underlay routing adapts to varying traffic)
• Question– Will the system reach a state with both low
latency and low network cost, as selfish routing and traffic engineering each tries to optimize their objective by adapting to the other process?
System & Network Reading Group
Specification of Vertical Interactions
• Interactive process between two players– Traffic engineering
• Given traffic matrix Tt, where Tt(s,d) denotes traffic from source s to destination d in time slot t
• Compute routing matrix Rt for the underlay• Objective: avoid overloading network
– Selfish routing• Given routing matrix Rt for the underlay• Produce new traffic matrix Tt
• Objective: minimize latency
System & Network Reading Group
One Round during Vertical Interaction
T(t) = Traffic matrix when routing matrix is R(t-1)
R(t) = OptimizedRoutingMatrix(T(t))Traffic engineering installs R(t) to
networkSelfish routing redistributes traffic to
form T(t+1)
System & Network Reading Group
Vertical Interaction with OSPF Optimizations
0.0E+00
5.0E+03
1.0E+04
1.5E+04
2.0E+04
2.5E+04
0 10 20 30 40 50
Round
Aver
age
late
ncy
(us)
overlay src: TE OSPF overlay src: hop-count compliant
020406080
100120140160180200
0 10 20 30 40 50
Round
Max
imum
link
util
izat
ion
(%)
overlay src: TE OSPF overlay src: hop-count compliant
OSPF route optimization interacts poorly with selfish routing
System & Network Reading Group
Vertical Interaction with MPLS Optimization
0.0E+002.0E+034.0E+036.0E+038.0E+031.0E+041.2E+041.4E+04
0 10 20 30 40 50
Round
Ave
rage
late
ncy
(us)
overlay src: TE MPLS overlay src: hop-count compliant
0
20
40
60
80
100
120
0 10 20 30 40 50
Round
Max
link
util
izat
ion
(%)
overlay src: TE OSPF overlay src: hop-count compliant
MPLS optimization interacts with selfish routing more effectively
System & Network Reading Group
Summary: Selfish Routing vs. Traffic Engineering
• OSPF route optimization interacts poorly with selfish routing
• MPLS interacts with selfish routing more effectively
• Despite the encouraging results from MPLS, several challenges exist– How to estimate traffic matrices accurately in
presence of adaptive selfish traffic?– Large optimization problems
System & Network Reading Group
Conclusion• Formulate and evaluate selfish overlay routing• Unlike the theoretical worst cases, selfish routing
in Internet-like environments yields close to optimal latency– The above result is true for both source
routing and overlay routing– Selfish routing can achieve good performance
without hurting the traffic that is using default routing
System & Network Reading Group
Conclusion• Mismatch between selfish routing and traffic
engineering– Different objectives
• Selfish routing: minimize e2e delay• Traffic engineering: aim to balance load
– Selfish routing reduces latency at the cost of increased congestion
– The adaptive nature of selfish routing makes traffic demands less predictable and reduces the effectiveness of traffic engineering
System & Network Reading Group
Future Work• Study impacts of multi-AS nature of the
Internet• Study dynamics of selfish routing (i.e.,
how traffic equilibria are reached?)• Improve the interactions between selfish
routing and traffic engineering• Study other selfish routing objectives
(e.g., loss and throughput)