On Selfish Routing in Internet-like Environments

30 October 2013

1

On Selfish Routing in Internet
-
like
Environments


paper by

Lili Qiu, Yang Richard Yang, Yin Zhang, Scott Shenker



presentation by

Ed Spitznagel

30 October 2013

2

Selfish Routing



•
Selfish routing: allowing hosts to choose routes
themselves (e.g. source routing) or use overlay
routing networks (e.g. Detour, RON)

–
routing decisions designed to optimize host
-
based or
overlay
-
based metrics, rather than system
-
wide metrics

30 October 2013

3

Selfish Routing


•
Theoretical results have shown that selfish routing
can result in “suboptimal system behaviour”

–
“price of anarchy” (latency of selfish routing vs.
globally optimal) can be unbounded for general latency
functions [Roughgardern and Tardos]


•
Does this happen in Internet
-
like environments?

30 October 2013

4

Outline


•
Models and Methodology

•
Selfish Source Routing

•
Selfish Overlay Routing

•
Selfish Routing vs. Traffic Engineering

•
Conclusion

30 October 2013

5

Network Model

•
Physical network: directed graph
G = (V, E)

where latency
of each link is a function of its load

•
Demands: Network traffic is partitioned into traffic
demands

•
Overlays: an overlay consists of overlay nodes, directed
overlay links, and a set of demands originated from the
overlay nodes.

–
paper considers the fully connected overlay topology only

•
Users: originators of traffic. Their objective is to minimize
latency.

30 October 2013

6

Network Model


•
Route Controller: controls the network
-
level
routing.

–
OSPF routing: three methods for weight assignment are
used:


Hop
-
count


Random
-
weight


Optimized
-
compliant: weights are set to minimize network
cost, assuming all traffic is compliant.

–
MPLS routing: uses multi
-
commodity flow routing

30 October 2013

7

Routing and Traffic Equilibria


•
Evaluate each selfish routing scheme by
computing its performance at traffic equilibria

•
Game
-
theoretic approach: equilibrium is a state
where no user can improve his traffic’s latency

•
Computed directly by algorithms in the appendix
(faster than using simulation techniques)

30 October 2013

8

Routing

•
Routing on the physical network

–
Source Routing: results in selfish routing of the type most often
studied

–
Optimal Routing: routing that achieves minimal average latency
over all traffic demands

•
Overlay routing

–
Overlay Source Routing: selfish routing through overlay nodes

–
Overlay Optimal Routing: routing through the overlay that
achieves minimal average latency

•
Compliant routing: follow the routes determined by the
network
-
level routing protocol

30 October 2013

9

Network Topologies

•
A real tier
-
1 ISP (referred to as
ISPTopo
) with ~100
backbone routers connected via OC48 and OC192 links

•
Other ISP topologies as mapped by Rocketfuel:







•
Random power
-
law topologies generated using BRITE

30 October 2013

10

Traffic demands

•
Real traffic demands: measured on
ISPTopo

for
three randomly chosen hours in November 2002

•
Synthetic traffic demands: randomly map POPs in
ISPTopo

to non
-
leaf nodes in the Rocketfuel
topology.

•
Load scale factor: scale up demands so that, when
all traffic is compliant and routed based on
shortest hop
-
count
, the max link utilization is
100∙
F
%, where
F

is the load scale factor

30 October 2013

11

Link Latency Functions

•
Several link latency functions are used:






•
“To avoid the discontinuity when the load approaches
capacity, we approximate the M/M/1 or M/D/1 function
with a linear function beyond 99% utilization”

–
(not sure if this makes sense)

30 October 2013

12

Performance Metrics

•
Average latency

•
Maximum link utilization

•
“Network cost” metric:

–
cost for a link with load
x

and capacity
c

is:





–
cost of a network is the sum of its links’ costs

30 October 2013

13

Outline


•
Models and Methodology

•
Selfish Source Routing

•
Selfish Overlay Routing

•
Selfish Routing vs. Traffic Engineering

•
Conclusion

30 October 2013

14

Selfish Source Routing



•
Are Internet
-
like environments among the worst cases?

30 October 2013

15

Selfish Source Routing

•
Effect of network load

on average latency:

30 October 2013

16

Selfish Source Routing

•
User latency for all topologies, using M/M/1 latency
function

30 October 2013

17

Selfish Source Routing

•
User latency for
ISPTopo
, varying the latency function

30 October 2013

18

System
-
wide cost for selfish source routing

•
Effect of network load

on max link utilization:

30 October 2013

19

System
-
wide cost for selfish source routing

•
Effects of network topologies:







•
selfish routing: highest max link utilization, highest network cost

•
selfish routing may make a network more prone to overload,
especially when failures occur

30 October 2013

20

System
-
wide cost for selfish source routing

•
Effects of latency functions:







–
results for different latency functions are qualitatively
the same

30 October 2013

21

Outline


•
Models and Methodology

•
Selfish Source Routing

•
Selfish Overlay Routing

•
Selfish Routing vs. Traffic Engineering

•
Conclusion

30 October 2013

22

When every node is in the overlay

•
ISPTopo, using OSPF for network
-
level routing

•
three choices of OSPF weight assignments

•
3 of the 4 curves overlap

30 October 2013

23

When every node is in the overlay

•
Again, but with other topologies

•
Selfish routing achieves latency close to optimal (and
usually better than compliant routing)

•
...but at the cost of higher maximum link utilization


30 October 2013

24

When not every node is in the overlay

•
ISPTopo
, overlay covers edge nodes only

•
curves almost completely overlap

–
Internet backbone is fairly well
-
connected and well
-
provisioned

30 October 2013

25

When not every node is in the overlay


•
Random partial coverage: latency similar to full coverage;
full coverage has higher max link utilization

30 October 2013

26

Interactions Among Competing Overlays

•
Competing demands at the same nodes, using
same underlying network

•
Foreground and background traffic at 50% each



Different topologies

Different methods for computing
OSPF weights

30 October 2013

27

Can Many Overlays Coexist Well?

•
Results for
ISPTopo
, with various multiple
-
overlay configurations:


30 October 2013

28

Outline


•
Models and Methodology

•
Selfish Source Routing

•
Selfish Overlay Routing

•
Selfish Routing vs. Traffic Engineering

•
Conclusion

30 October 2013

29

Selfish Routing vs. Traffic Engineering

•
Traffic engineering: adjust physical routing based
on network traffic patterns (traffic matrix)

•
Selfish routing will then adjust the traffic matrix in
response to that…

•
… which causes the traffic engineering to produce
a new routing matrix

•
And the process repeats…

30 October 2013

30

Selfish Routing and OSPF optimizer

•
Optimizer chooses link weights for OSPF







•
Selfish routing + traffic engineering: considerably
worse performance than either one alone

30 October 2013

31

Selfish Routing and MPLS optimizer

•
Better, perhaps due to MPLS’s finer
-
grained control (can adjust
routing matrix without reducing available resources, unlike OSPF)

•
Almost as good as using only selfish routing or only TE

•
MPLS TE requires solving very large linear programming problem

30 October 2013

32

Conclusion

•
Unlike the worst
-
case scenarios, selfish routing in
Internet
-
like environments achieves close to optimal

–
often at the cost of overloading certain links

–
can become worse due to interaction with traffic engineering

•
Future work

–
investigate how the multi
-
AS nature of Internet affects routing
performance

–
dynamics of selfish routing (how equilibria are reached)

–
better understanding of interactions between traffic engineering
and selfish routing

–
study of selfish routing with other metrics (loss, throughput)

30 October 2013

33

End of Presentation


•
Questions?