Failure Inferencing based Fast Rerouting for Handling Transient Link and Node Failures

1
Failure Inferencing based Fast Rerouting for
Handling Transient Link and Node Failures

• Zifei Zhong
• University of South Carolina, Columbia
• Joint work with
• Srihari Nelakuditi, Junling Wang
• University of South Carolina, Columbia
• Sanghwan Lee, Yinzhe Yu
• University of Minnesota, Minneapolis
• Chen-Nee Chuah
• University of California, Davis

2
Outline

• Motivation
• Wired Network Failure Characteristics
• Service Availability Expectations
• Limitations of Current Routing Schemes
• Failure Inferencing based Fast Rerouting (FIFR)
• FIFRL for single link failures IWQoS03,
  Infocom04
• FIFRN for single link node failures
• FIFRN Evaluation Summary

3
Failure Characteristics

• Study by Sprintlabs on failures in an IP backbone
  Infocom04
• Failures are fairly common, occur almost everyday
• Maintenance, faulty interfaces, router crashes,
  fiber cuts, misconfigurations
• Majority of the unplanned failures are transient
• 46 lt 1 minute, 86 lt 10 minutes
• Majority of them are single link/node failures
• 85 of unplanned failures affect a single
  link/router
• Focus transient single link and node failures

4
Availability Expectations

• Increased expectation from the Internet
• Wish for five nines IP network
• Disruption-sensitive applications
• Voice over IP
• Break gt 60 ms in voice traffic is noticeable
• E-commerce
• Even small downtime affects business and
  reputation
• High capacity links
• Short outage, large impact
• Link down for 10 seconds ? 3 million packets lost
• Goal high availability despite transient failures

5
Current Schemes for Failure Resiliency

• Traditional link state routing protocols
  (OSPF/ISIS)
• React to failures with global rerouting
• Global link state updates routing table
  recomputations
• Trade-off between stability and continuity
• Convergence delay to resume forwarding after a
  link failure
• MPLS based recovery
• Local rerouting along a preconfigured LSP
• Label stacking enables tunneling of affected LSPs
  thru protection LSP
• Requires a shift to label switching paradigm
• Label swapping instead of destination based
  forwarding

6
Failure Inferencing based Fast Rerouting

• Provides fast local loop-free rerouting
• Without explicit failure notification
• Prepares for failures
• Failure inferencing
• Infer failures based on packets incoming
  interface
• Interface-specific forwarding
• Next hop based on destination and incoming
  interface
• Local rerouting upon adjacent link/node failures
• Suppress link state update and trigger local
  rerouting

7
Illustration No Failure Scenario
Route (1 to 6) 1-gt2-gt5-gt6
8
Illustration Local Rerouting without FIFR
Route (1 to 6) 1-gt2-gt1-gt2-gt
-gt (loop!!)
9
Illustration Local Rerouting with FIFRL
Route (1 to 6) 1-gt2-gt1-gt3-gt5-gt6
10
Illustration FIFRL with Node Failures
Route (1 to 6) 1-gt2-gt1-gt3-gt1-gt-gt (loop!!)
11
Inferencing for Node Failures

• Infer node failures instead of link failures?
• Yes, but how?
• Can it still guarantee loop-freedom?
• Yes.

12
Illustration Local Rerouting with FIFRN
Route (1 to 6) 1-gt2-gt1-gt4-gt6
13
Forwarding Table Computation

• Assumptions
• Links are bidirectional with equal weight
• At most a single node failure is suppressed
• Infer failed nodes from packets arrival at an
  interface
• set of key nodes whose failure causes
  packet to d arrive at i from j
• A node u is included in the set of key nodes if
• with u, j is a next hop from i to d
• without u, edge j?i is along the shortest path
  from the upstream of u to d
• Avoid all key nodes in choosing packets next hop
• set of next hops to d from i when packet
  arrives at i from j,
• Forwarding tables are pre-computed

14
Illustration Key Nodes Computation
15
Loop-freedom of FIFRN

• Forwarding around the failed node
• When no more than one node failure is suppressed,
  FIFRN can find a loop-free path to a destination
  if one such path exists

16
Performance Evaluation

• Stretch
• Stretch vs. nodes

Varying number of nodes with degree 6
17
Performance Evaluation

• Stretch vs. degrees

Varying number of degree with 200 nodes
18
FIFRN Summary

• Fast reroute for IP networks without MPLS
• Without explicit failure notification
• Protection against any singe node failures
• Can actually handle both single link and node
  failures
• Better stability and availability than OSPF
• Particularly when failures are frequent and
  transient
• Minimal changes to current network infrastructure
• Only need to replace SPF algorithm with FIFR
  algorithm

19
Questions, please! ?