15-744: Computer Networking

Download Report

Transcript 15-744: Computer Networking

15-744: Computer Networking
L-4 Network Topology
Today’s Lecture
•
•
•
•
•
Structural generators
Power laws
HOT graphs
Graph generators
Assigned reading
• On Power-Law Relationships of the Internet
Topology
• A First Principles Approach to Understanding
the Internet’s Router-level Topology
2
Outline
• Motivation/Background
• Power Laws
• Optimization Models
• Graph Generation
3
Why study topology?
• Correctness of network protocols typically
independent of topology
• Performance of networks critically dependent on
topology
• e.g., convergence of route information
• “time-travel”
• Internet impossible to replicate
• Modeling of topology needed to generate test
topologies
• Incentives/Mechanisms to create topologies
• Identifying good topologies and mechanism design
for them
4
Internet topologies
AT&T
MCI
Router level
AT&T
MCI
SPRINT
SPRINT
Autonomous System (AS) level
5
More on topologies..
• Router level topologies reflect physical connectivity
between nodes
• Inferred from tools like traceroute or well known public
measurement projects like Mercator and Skitter
• AS graph reflects a peering relationship between two
providers/clients
• Inferred from inter-domain routers that run BGP and publlic
projects like Oregon Route Views
• Inferring both is difficult, and often inaccurate
6
Hub-and-Spoke Topology
• Single hub node
• Common in enterprise networks
• Main location and satellite sites
• Simple design and trivial routing
• Problems
•
•
•
•
Single point of failure
Bandwidth limitations
High delay between sites
Costs to backhaul to hub
7
Simple Alternatives to Hub-and-Spoke
• Dual hub-and-spoke
• Higher reliability
• Higher cost
• Good building block
• Levels of hierarchy
• Reduce backhaul cost
• Aggregate the
bandwidth
• Shorter site-to-site
delay
…
8
Abilene Internet2 Backbone
9
Intermountain
GigaPoP
Front Range
GigaPoP
Indiana GigaPoP
U. Louisville
Great Plains
Merit
OARNET
OneNet
Qwest Labs
Arizona St.
Northern Lights
U. Memphis
WiscREN
NCSA
StarLight
U. Arizona
Iowa St.
MREN
Oregon
GigaPoP
Pacific
Northwest
GigaPoP
NYSERNet
Pacific
Wave
UNM
Denver
Kansas
City
WPI
Indianapolis
Northern
Crossroads
Chicago
Seattle
SINet
U. Hawaii
New York
ESnet
AMES NGIX
GEANT
WIDE
Sunnyvale
CENIC
SURFNet
Wash
D.C.
Rutgers U.
Los Angeles
UniNet
TransPAC/APAN
Houston
North Texas
GigaPoP
SFGP/
AMPATH
Texas
GigaPoP
Miss State
GigaPoP
(as of December 16, 2003)
0.1-0.5 Gbps
0.5-1.0 Gbps
1.0-5.0 Gbps
5.0-10.0 Gbps
UT Austin
UT-SW
Med Ctr.
Atlanta
SOX
Texas Tech
Abilene Backbone
Physical Connectivity
MANLAN
LaNet
Florida A&M
Tulane U.
U. So. Florida
MAGPI
PSC
DARPA
BossNet
UMD NGIX
Mid-Atlantic
Crossroads
Drexel U.
U. Florida
U. Delaware
NCNI/MCNC
10
Points-of-Presence (PoPs)
• Inter-PoP links
• Long distances
• High bandwidth
Inter-PoP
Intra-PoP
• Intra-PoP links
• Short cables between
racks or floors
• Aggregated bandwidth
• Links to other
networks
Other networks
• Wide range of media
and bandwidth
11
Deciding Where to Locate Nodes and Links
• Placing Points-of-Presence (PoPs)
•
•
•
•
Large population of potential customers
Other providers or exchange points
Cost and availability of real-estate
Mostly in major metropolitan areas
• Placing links between PoPs
• Already fiber in the ground
• Needed to limit propagation delay
• Needed to handle the traffic load
12
Trends in Topology Modeling
Observation
Modeling Approach
• Long-range links are expensive
• Random graph (Waxman88)
• Real networks are not random,
but have obvious hierarchy
• Structural models (GT-ITM
Calvert/Zegura, 1996)
• Internet topologies exhibit
power law degree distributions
(Faloutsos et al., 1999)
• Degree-based models replicate
power-law degree sequences
• Physical networks have hard
technological (and economic)
constraints.
• Optimization-driven models
topologies consistent with design
tradeoffs of network engineers
13
Waxman model (Waxman 1988)
• Router level model
• Nodes placed at random
in 2-d space with
dimension L
• Probability of edge (u,v):
u
d(u,v)
v
• ae^{-d/(bL)}, where d is
Euclidean distance (u,v), a
and b are constants
• Models locality
14
Real world topologies
• Real networks exhibit
•
•
•
•
Hierarchical structure
Specialized nodes (transit, stub..)
Connectivity requirements
Redundancy
• Characteristics incorporated into the
Georgia Tech Internetwork Topology
Models (GT-ITM) simulator (E. Zegura,
K.Calvert and M.J. Donahoo, 1995)
15
Transit-stub model (Zegura 1997)
• Router level model
• Transit domains
• placed in 2-d space
• populated with routers
• connected to each other
• Stub domains
• placed in 2-d space
• populated with routers
• connected to transit
domains
• Models hierarchy
16
So…are we done?
• No!
• In 1999, Faloutsos, Faloutsos and
Faloutsos published a paper, demonstrating
power law relationships in Internet graphs
• Specifically, the node degree distribution
exhibited power laws
That Changed Everything…..
17
Outline
• Motivation/Background
• Power Laws
• Optimization Models
• Graph Generation
18
Power laws in AS level topology
19
Power Laws and Internet Topology
Source: Faloutsos et al. (1999)
Rank R(d)
R(d) = P (D>d) x #nodes
Most nodes have few connections
Degree d
A few nodes have lots of connections
• Router-level graph & Autonomous System (AS) graph
• Led to active research in degree-based network models
GT-ITM abandoned..
• GT-ITM did not give power law degree
graphs
• New topology generators and explanation
for power law degrees were sought
• Focus of generators to match degree
distribution of observed graph
21
Inet (Jin 2000)
• Generate degree sequence
• Build spanning tree over nodes
with degree larger than 1,
using preferential connectivity
• randomly select node u not in
tree
• join u to existing node v with
probability d(v)/d(w)
• Connect degree 1 nodes using
preferential connectivity
• Add remaining edges using
preferential connectivity
22
Power law random graph (PLRG)
• Operations
•
•
•
•
assign degrees to nodes drawn from power law distribution
create kv copies of node v; kv degree of v.
randomly match nodes in pool
aggregate edges
2
1
1
may be disconnected, contain multiple edges, self-loops
• contains unique giant component for right choice of
parameters
23
Barabasi model: fixed exponent
• incremental growth
• initially, m0 nodes
• step: add new node i with m edges
• linear preferential attachment
• connect to node i with probability ki / ∑ kj
0.5
0.5
0.5
0.25
existing node
0.25
new node
may contain multi-edges, self-loops
24
Features of Degree-Based Models
Preferential Attachment
Expected Degree Sequence
• Degree sequence follows a power law (by
construction)
• High-degree nodes correspond to highly connected
central “hubs”, which are crucial to the system
• Achilles’ heel: robust to random failure, fragile to
specific attack
25
Does Internet graph have these properties?
• No…(There is no Memphis!)
• Emphasis on degree distribution - structure
ignored
• Real Internet very structured
• Evolution of graph is highly constrained
26
Problem With Power Law
• ... but they're descriptive models!
• No correct physical explanation, need an
understanding of:
• the driving force behind deployment
• the driving force behind growth
27
Outline
• Motivation/Background
• Power Laws
• Optimization Models
• Graph Generation
28
Li et al.
• Consider the explicit design of the Internet
• Annotated network graphs (capacity,
bandwidth)
• Technological and economic limitations
• Network performance
• Seek a theory for Internet topology that is
explanatory and not merely descriptive.
• Explain high variability in network connectivity
• Ability to match large scale statistics (e.g.
power laws) is only secondary evidence
29
Router Technology Constraint
3
10
Cisco 12416 GSR, circa 2002
Total Bandwidth
high BW low
degree
high degree
low BW
2
Bandwidth (Gbps)
10
1
10
Bandwidth per Degree
15 x 10 GE
15 x 3 x 1 GE
0
10
15 x 4 x OC12
15 x 8 FE
Technology constraint
10
-1
10
0
10
1
Degree
10
2
30
Aggregate Router Feasibility
cisco 12416
1000000
core technologies
approximate
aggregate
feasible region
Total Router BW (Mbps)
100000
cisco 12410
cisco 12406
10000
cisco 12404
1000
cisco 7500
100
cisco 7200
10
older/cheaper
technologies
1
linksys 4-port router
uBR7246 cmts
(cable)
cisco 6260 dslam
(DSL)
edge technologies
0.1
0.01
1
10
100
1000
10000
cisco AS5850
(dialup)
degree
Source: Cisco Product Catalog, June 2002
31
Variability in End-User Bandwidths
Connection Speed (Mbps)
1e4
Ethernet
1-10Gbps
1e3
1e2
high
performance
computing
a few users have very
Ethernet
high speed
10-100Mbps
connections
1e1
academic
and corporate
residential and
small business
Broadband
Cable/DSL
~500Kbps
1
1e-1
most users
have low speed
connections
1e-2
1
1e2
1e4
Rank (number of users)
Dial-up
~56Kbps
1e6
1e8
32
Heuristically Optimal Topology
Mesh-like core of fast, low degree routers
Cores
High degree nodes
are atEdges
the edges.
Hosts
33
Comparison Metric: Network Performance
Given realistic technology constraints on routers, how well
is the network able to carry traffic?
Step 1: Constrain to
be feasible
Step 2: Compute traffic demand
Bj
1000000
Bandwidth (Mbps)
xij  Bi B j
xij
100000
10000
1000
100
Bi
Abstracted
Technologically
Feasible Region
Step 3: Compute max flow
max  xij  max  Bi B j

10
degree
1
10
100
1000
i, j
s.t.
i, j
x
i , j:krij
ij
 Bk , k
34
Likelihood-Related Metric
Define the metric
L( g ) 
d d
i
i, j
connected
j
(di = degree of
node i)
• Easily computed for any graph
• Depends on the structure of the graph, not the generation
mechanism
• Measures how “hub-like” the network core is
• For graphs resulting from probabilistic construction (e.g.
PLRG/GRG),
LogLikelihood (LLH)  L(g)
• Interpretation: How likely is a particular graph (having given
node degree distribution) to be constructed?
35
Abilene-inspired
HOT
PA
Sub-optimal
PLRG/GRG
P(g)
Perfomance (bps)
12
10
11
10
Lmax
l(g) = 1
P(g) = 1.08 x 1010
10
10
0
0.2
0.4
0.6
l(g) = Relative Likelihood
0.8
1
36
Structure Determines Performance
HOT
P(g) = 1.13 x 1012
PA
P(g) = 1.19 x 1010
PLRG/GRG
P(g) = 1.64 x 1010
37
Summary Network Topology
• Faloutsos3 [SIGCOMM99] on Internet topology
• Observed many “power laws” in the Internet structure
• Router level connections, AS-level connections, neighborhood sizes
• Power law observation refuted later, Lakhina [INFOCOM00]
• Inspired many degree-based topology generators
• Compared properties of generated graphs with those of measured
graphs to validate generator
• What is wrong with these topologies? Li et al [SIGCOMM04]
• Many graphs with similar distribution have different properties
• Random graph generation models don’t have network-intrinsic
meaning
• Should look at fundamental trade-offs to understand topology
• Technology constraints and economic trade-offs
• Graphs arising out of such generation better explain topology and its
properties, but are unlikely to be generated by random processes!
38
Outline
• Motivation/Background
• Power Laws
• Optimization Models
• Graph Generation
39
Graph Generation
• Many important topology metrics
•
•
•
•
Spectrum
Distance distribution
Degree distribution
Clustering…
• No way to reproduce most of the important
metrics
• No guarantee there will not be any
other/new metric found important
40
dK-series approach
• Look at inter-dependencies among topology
characteristics
• See if by reproducing most basic, simple,
but not necessarily practically relevant
characteristics, we can also reproduce
(capture) all other characteristics, including
practically important
• Try to find the one(s) defining all others
• dK-distributions are degree correlations within
simple connected graphs of size d
0K
Average degree <k>
1K
Degree distribution P(k)
2K
Joint degree distribution P(k1,k2)
3K
“Joint edge degree” distribution P(k1,k2,k3)
3K, more exactly
4K
Rewiring
50
Graph Reproduction
51
The elephant in the room…
• How good is the underlying data on which
these studies are based?
• E.g., sampling bias  traceroute of shortest
paths on random graph can produce powerlaw distribution [Lakhina03]
• Similar issues with AS-level view
• Router level data is very noisy
52
Better Measurements?
• Rocketfuel [sigcomm02]
• Better router alias resolution
• Detailed maps based on multiple viewpoints
• Reverse traceroute [nsdi10]
• Improves ability to view peer-to-peer links
• RouteViews and BGP collection efforts
53
Next Lecture: Congestion Control
• No lecture Monday
• Congestion Control
• Assigned Reading
• [Jacobson and Karels] Congestion Avoidance
and Control
• [TFRC] Equation-Based Congestion Control for
Unicast Applications (sections 1-3)
54