Transcript R i - Computer Science and Engineering

Toward a Fair and Robust
Internet Backbone Routing
Jerry Chou, UC San Diego
Bill Lin, UC San Diego
Shubho Sen, AT&T Labs
Oliver Spatscheck, AT&T Labs
Outline
• Problem & Motivation
• Current Solutions
• Experimental Results
• Discussion & Conclusion
UC Symposium, August 4, 2006 @ AT&T
2
Same Old Problem
• Robustness
• Efficiency
– Support variable
traffic
– Fault tolerance
• Fairness
– High throughput
– Minimize
network
resources
– Fairness policy
(max-min/
proportion)
Internet Backbone Network
R
ISP
R
R
ISP
R
R
R
UC Symposium, August 4, 2006 @ AT&T
R
R
R
R
R
R
R
ISP
R
Customer
Network
3
Getting Worse…
• Traffic is rapidly changing
– New applications: P2P applications
→ Unpredictable traffic
– New
→ Growing traffic
Lowcustomers
efficiency
– New technology: Broadband/optical networks to homes
No performance guarantee
→ move bottleneck toward core network
Weak robustness
• Routing is still OSPF
– Basic approach
• Limited computational power on routers
• Distributed environment
– Topology dependent
• Traffic engineered network
• Over provision
UC Symposium, August 4, 2006 @ AT&T
4
Goals
• Classify routing algorithm
• Study the design decision of routing
• Evaluate and compare routing algorithms
• Design a better routing algorithm
UC Symposium, August 4, 2006 @ AT&T
5
Outline
• Problem & Motivation
• Current Solutions
• Experimental Results
• Discussion & Conclusion
UC Symposium, August 4, 2006 @ AT&T
6
Two Problem Formulations
• Optimize for a given traffic profile
– Use state-of-the-art network measurement techniques
to measure traffic conditions
– Find at run-time optimal path load-distribution for
measured traffic profile
– No need to change previous path load-distribution if it
can handle current traffic condition – only change
when necessary
• Optimize for worst-case, assuming all
“admissible” traffic profiles are possible
– Determine optimal path load-distribution under this
worst-case model once offline
UC Symposium, August 4, 2006 @ AT&T
7
Optimizing for Specific Traffic Profile
• Maximum Concurrent Flow Model
– Use a centralized control plane model – e.g. AT&T’s
RCP proposal
– Periodic traffic measurements and topology changing
reported back to central controller
– For given traffic matrix T, find optimal path loaddistribution on topology graph G.
R
R
R
R
R
R
R
R
UC Symposium, August 4, 2006 @ AT&T
8
Optimize for Worst-Case Traffic
• 2Phase on Hose Model:
– Ri = max ingress traffic at node i,
– Ci = max egress traffic at node i
Ri
Ci
– Traffic matrix T is admissible iff
• Sj tij <= Ri , Sj tji <= Ci , for all i
• Set of all such traffic matrices denoted by t ( R , C )
– Routing algorithm
• Phase1 splits traffic to all nodes
• Phase2 redirects traffic to destination
UC Symposium, August 4, 2006 @ AT&T
9
Outline
• Problem & Motivation
• Current Solutions
• Experimental Results
• Discussion & Conclusion
UC Symposium, August 4, 2006 @ AT&T
10
Routing Algorithm Explore
• Open Shortest Path First (OSPF) [1]
– Route all flows with the same source and destination on
a single shortest path
• Load-balanced OSPF (LOSPF) [1]
– Evenly distribute the load of a flow in packet level on all
shortest paths between its source and destination
• Maximum Concurrent Flow Routing [2]
– Route flow based on maximum concurrent flow problem
formulation for a given traffic matrix and topology
• 2PHASE [3]
– Phase1 splits flows to all nodes with a fix ratio
determined by the access capacity
– Phase2 redirects flows to their destination nodes
UC Symposium, August 4, 2006 @ AT&T
11
Initial Experimental Results
• Testbed:
– Large ISP network backbone
changes
very day
– •Topology
Degree-based
Topology
Generator [4]
– •Traffic
Random
Graph Generator
changes
every hour[5]
• Matrix
– Link Load:
Lambda = 2
traffic load on a link
link capacity
50%
40%
1
– Lambda:
max link load among all links
– Throughput:
Total traffic comes out the network
Total traffic goes in the network
– Cost: Additional capacity required to support a traffic
UC Symposium, August 4, 2006 @ AT&T
12
Lambda
Lambda
3
MCF
LOSPF
2PHASE
OSPF
2.5
2
1.5
1
0
4
8 12 16 20 24 28 32 36 40 44 48
Hour (from 06/28/06 00am)
•
•
•
•
In general: MCF > LOSPF > 2Phase > OSPF
Topology is traffic engineered for OSPF
2Phase guarantees ONLY 10% of the worse case traffic
2Phase has higher link load due to long routing path
UC Symposium, August 4, 2006 @ AT&T
13
Throughput (%)
Throughput
100
95
90
85
80
75
70
65
60
MCF
LOSPF
2PHASE
OSPF
1
•
•
•
•
•
3
5
Load Factor
We increase the demand amount per flow by load factor
The bar specifies the average throughput in 2days period
Error bar is the max and min throughput in 2days period
2Phase has the largest variation
MCF always maintains high throughput
UC Symposium, August 4, 2006 @ AT&T
14
Future Plan & Discussion
• For MCF:
– How often to perform optimization?
– What is the cost to reconfigure the routing?
• For 2Phase:
– How to formulate a more strict worse case?
• By access capacity, core link capacity, demand
amount or something else?
• For OSPF:
– What’s the general topology for networks?
– What’s the impact of topology to those routing
algorithm?
UC Symposium, August 4, 2006 @ AT&T
15
Reference
[1] “Configuring OSPF”, Cisco Systems Prgoduct Doc.
[2] G. Karakostas, “Faster Approximation Schemes for
Fractional Multicommodity Flow Problems”.
[3] M. Kodialam, T. V. Lakshman, J. B. Orlin and S.
Sengupta, “A Versatile Scheme for Routing Highly
Variable Traffic in Service Overlays and IP Backbones”.
[4] C. Jin, Q. Chen, and S. Jamin, “Inet: Internet Topology
Generator”, Technical Report CSE-TR443 -00,
Department of EECS, University of Michigan, 2000.
[5] B. M. Waxman. Routing of multipoint connections. IEEE
Journal of Selected Areas in Communications, pages
1617--1622, 1988.
UC Symposium, August 4, 2006 @ AT&T
16
Backup Slides
UC Symposium, August 4, 2006 @ AT&T
17
Two Phase Routing
• Approach
– Phase1 splits traffic to all nodes by
a fix ratio
– The ratio is determined by the
access capacity
– Phase2 delivers demand to
destination
Logic Network:
R1
• Support all valid traffic matrix
– High fault tolerance
• Cons:
– Long path: 2*network diameter
=>Waste network capacity
– Access capacity >> link capacity
UC Symposium, August 4, 2006 @ AT&T
l
l=2(r/N)
R2
l
R3
r
• Pros:
– Independent of traffic
r
r
Logic Routing Path:
r
r/N
R1
R1
r
r
R2
R3
R2
r/N
R1
R2
R3
R3
Phase1 Phase2
18