On Multi-Path Routing

1
On Multi-Path Routing

• Aditya Akella
• 03/25/02

2
What is Multi-Path Routing?

• Dynamically route traffic
• Multiple paths to a destination
• Path taken dependant on the relative load on
  candidate paths at flow-arrival time
• Help flows circumvent congested links
• Load balancing, improved response time

3
Static routing

• For example, OSPF
• After route update, pick path P with least cost
• All flows between the same source and destination
  will follow P until the next update picks a
  different path P
• Best candidate path information is stale for the
  update period

4
Multi-path routing vs. Static routing

• Best candidate path of a flow is a function of
  the network state when the flow arrives
• Avoids staleness, in some sense
• Keeping accurate network state at every instant
  of time very tough
• Multi-path routing uses estimated network state
• Update estimated state at regular intervals

5
Past work and state-of-the-art

• Plenty of related work (but, will mainly
  discuss Shaikh and Rexfords work)
• Perform load-sensitive routing of long-lived IP
  flows
• Statically route short lived flows
• Applicable to backbone networks of ISPs
• First hop router of a flow computes the path for
  the flow based on knowledge of the entire network
  (the backbone)
• Link State routing
• Intra-AS solution

6
A different take on the problem

• Multi-Path routing of IP flows
• Existing solutions designed to work within the
  core of an AS
• Intra-AS
• Aimed at achieving optimum provisioning of
  resources within an AS
• But the core is relatively under-utilized
• Multi-path routing might only help make
  under-utilization unifrom across an ISP
• Greater congestion probably occurs at the edges
  of an AS
• Queueing drops at Peering points
• Edge of the network where stub networks attach
• Might help more to do multi-path at inter-AS level

7
So ?

• Existing solutions should be extended to work at
  a coarser level
• Both Intra-AS and Inter-AS
• How about designing a solution that works at the
  Inter-AS level?
• Focus of this work
• Can work in combination with an Intra-AS solution

8
In this talk

• Explore the utility of employing multi-path
  routing at AS-Level
• What issues hinder the usefulness?
• How can they be analyzed?
• I will not present an actual mechanism

9
How useful can this be?
(1) BGP

• Potential multiple paths are hidden by BGP
• Policy
• E.g., Peer will not provide transit service
• Path Vector as opposed link state
• Only path length information
• Need path cost information
• Neighbor relations do not reveal number of paths
• Neighbors could peer at many places

10
BGP contd.

• How useful is Multi-Path Routing?
• How much potential for Multi-Path Routing exists,
  with BGP in its current form?
• How much potential can we add and how?
• Some factors just cannot be helped (some
  policies)
• What other factors can we get around?

11
How useful can this be?
(2) The Costs

• The cost function
• What is it, exactly?
• How can it be computed in a distributed manner?
• At what frequency to compute, update routes?
• Etc

12
How useful can this be?
(3) Selfish Routing

• What if everybody employed Multi-Path Routing?
• Each flow is now a selfish agent
• Each flow tries to maximize its observed
  performance
• What would be the marginal utility per flow when
  all flows are selfish?
• Nash Equilibrium
• Cannot be better than the optimal case --
• When flows route packets in such a manner as
  to optimize a common global metric social
  equilibrium
• Game-Theoretic Analysis

13
Game Theory

• Will not bore you with the details. Here is an
  outline
• We have a game-theoretic model for multi-path
  routing
• Have a way of showing how bad Nash equilibrium is
  compared to social equilibrium
• Have a link pricing mechanism that can be
  employed to bridge the gap between Nash and
  social equilibria
• Key idea cheat by propagating link costs
  different from the true costs
• Greedy flows will try to minimize cost
• Design cost in such a way that greedy flows
  actually end up minimizing average cost (social
  optimum)

14
Thats about it!

• Questions/Comments?