Cost-Performance Tradeoffs in MPLS and IP Routing

Selma YilmazIbrahim Matta

Boston University

Motivation

Conventional IP Routing

Static shortest-path destination based only routing Plus

– Single state per destination in the forwarding tableMinus

– Leads to unbalanced traffic distribution

The Fish Example R2-R3-R4 may get over-utilized R2-R6-R7-R4 may stay under-

utilized

Find ways to make better utilization of resources by making use of alternate paths

30-80% of the cases there is an alternate path with significantly superior quality i.e. loss rate, bandwidth, RTT [Savage:Sigcomm99]

Classification of Routing Solutions

• Best-effort solutions Ex: Per-packet dynamic routing

• QoS routing solutions Ex: Widest-shortest path (WSP)

• QoS and traffic aware solutions

– Location of ingress-egress pairs Ex: Minimum Interference Routing (MIRA) [Kar:Infocom00]

– Traffic matrix– Both location of ingress-egress pairs and traffic matrix Ex: Profile-based routing (PBR) [Suri:01]

How far are these solutions from optimal?

Available Resources QoS Requirements Traffic DemandsBest-effort

QoS Routing

QoS and Traffic Aware Routing

+

+

+

+

+ +

Cost – Time Complexity – Space Complexity

Performance Measures– Bandwidth Acceptance Ratio

Total bandwidth accepted/Total bandwidth requested– Utility Per-packet: Portion of flow that is accepted Per-flow: 0/1

– Maximum Link Utilization

•Per-flow (MPLS kind)•Guaranteed bandwidth

•Stateless•Best effort

Increasing Cost

Increasing Performance ?

Per-packet Dynamic Routing WSP MIRA PBR

Review of Evaluated Algorithms

Per-packet Dynamic Routing

Properties– Avoids congested links – Computationally simple – Stateless

Difficulties

– Link states change at packet level– Impractical to generate link state updates at packet level– Larger link state update periods may cause oscillations

Widest-Shortest Path Routing

– Choose feasible min-hop path

Break ties by picking the widest

– limit resource consumption: shortest paths

• improves performance under heavy load

– balance load: widest paths

• increases utilization and long term performance

– Per-flow state is maintained

– Run time complexity is same as Dijkstra’s shortest path algorithm

Minimum Interference Routing (MIRA)

Goal: Increase utilization and long term performance of QoS routing by being aware of location of ingress-egress pairs

Idea: Among feasible paths, pick the one that interferes the least with future requests

• Link costs are assigned based on criticality

• Shortest path routing

• Run time complexity is complexity of maxflow computation

• Per-flow state

11

6

107

4321 5

98

S1

S2

S3

D1

D2

D3

Profile-based Routing (PBR)

Goal: Increase utilization and long term performance of QoS routing by being aware of location of ingress-egress pairs and traffic matrix

Traffic Profile (classID, si, di, Bi): Aggregate expected traffic between

ingress si -egress di for a class classID.

Idea:

– Using offline phase to compute pre-allocated capacities for each traffic class

– Routing during online phase within these pre-allocated capacities

Profile-based Routing (PBR)

Off-line (pre-processing) phase

• Compute an optimal distribution of profiles by solving multicommodity flow problem

xi(e) amount of commodity i routed through edge e

• Each profile is a commodity • Excess edges are added to always have feasible solution

• Flows are forced to be routed through original edges as much as

possible

D2cost=infinity

cost=infinityS1 D1

cost=1

S2

Simulations

• Algorithms: Dynamic Per-packet Routing, WSP, MIRA, PBR

Dynamic per-packet routing

– Multicommodity flow problem is solved at each flow arrival/departure

xi(e) amount of commodity i routed through edge e

– Each active flow is a commodity and allowed to split

– Excess edges are used to always have feasible solution

– During its lifetime, an individual flow can

• be split

• get different bandwidth values

• be assigned to different paths

Simulation Model and Performance Measures

Assumptions– Only one class between an ingress-egress pair– Bandwidth demands of flows that belong to the same class are

same– Flow arrivals from a class is according to Poisson process– Hold times are Pareto– Load between different ingress-egress pairs is same

Performance Measures– Bandwidth Acceptance Ratio

Total bandwidth accepted/ Total bandwidth requested– Utility Per-packet: Portion of flow that is accepted Per-flow: 0/1

– Maximum Link Utilization

ResultsRainbow Topology

Per-packet Dynamic Routing > WSP~MIRA > PBR

Results

Rainbow Topology

Profiles are (class1,S1,D1,2) and (class1,S2,D2,2)Accepted bandwidth with MIRA=WSP=n, PBR=0

1

2

3

5

6 7

8 9 10

11 12 13 14

1

1

11

S1

S2

D1D2

1

2

1

1

1 11

4

4 4 4

4

2

4

Results

• Per-packet dynamic routing packs load along shortest paths - increases maximum utilization

• PBR has lowest maximum utilization - it lets links stay underutilized

• WSP should be best at load balancing - not seen since there is no alternate paths

Regular Topology

Results

Conclusion

• Dynamic per-packet routing shows best, PBR shows worst performance

• Among per-flow routing algorithms, WSP shows good performance at low cost

• More information doesn’t mean more gain• Because of pre-allocation, statistical multiplexing is lost

Future Work

• Using extra information in the form of traffic matrix and ingress-egress pairs should lead to a better performance. Why it didn’t?– Take traffic variability into consideration– Don’t take pre-allocated capacities as hard limits

• For cost-performance tradeoffs – Solutions that are at the extreme ends of spectrum, i.e. no

state or per-flow state, not practical• Find good operating point where

– performance is good enough– cost is not too high

Ex: hybrid routing for traffic classes