Chapter%204%20Link-State%20Routing%20and%20Hierarchical%20Routing

1
Chapter 4Link-State Routing and Hierarchical
Routing

• Professor Rick Han
• University of Colorado at Boulder
• rhan_at_cs.colorado.edu

2
Announcements

• Handing back HW 1, solutions online
• Homework 2, due Feb. 26
• Midterm for the week of March 12
• Next, link-state routing, hierarchical routing

3
Recap of Previous Lecture

• Problems with Distance Vector Loops
• Bouncing Effect
• Bad news propagates slowly
• Counting to Infinity
• Split Horizon with Poison Reverse
• Dijkstra as alternative Shortest Path algorithm
  to distributed Bellman-Ford Equation
• Iteratively grow a shortest path spanning tree
  from the root outwards
• Link-state routing
• Dijkstra shortest path algorithm,
• Reliable flooding of LSPs to all nodes

4
Link State vs. Distance Vector

• Routing update size
• LS small, contain only neighbors link costs
• DV potentially long distance vectors (length N
  for N nodes in network)
• Routing update communication overhead
• LS flood to all nodes, overhead is O(NE), where
  N is nodes, and E is edges or links
• In DV, send distance vector only to neighbors

5
Link State vs. Distance Vector(2)

• Convergence speed
• DV at each iteration, send to neighbor and
  recalculate
• takes awhile to propagate changes to rest of
  network
• Iterations are periodic, hence slow faster when
  triggered
• LS faster dont need to recalculate LSPs
  before forwarding
• may be key reason that LS beat out DV in
  intra-domain routing

6
Link State vs. Distance Vector (3)

• Space requirements
• LS maintains entire topology in a link database
• DV maintains only neighbor state
• If each of N routers has K neighbors,
• LS O(NK) memory requirement
• DV O(NK) also

7
Link State vs. Distance Vector (4)

• Complexity (initializing from scratch)
• DV O(NKDiameter)
• for each of N rows in distance table, find min of
  (dikDkj) over K neighbors, iterate until get
  info via DVs of furthest nodes (a diameter away)
• If sorted list kept, reduce complexity to
  O(Nlog(K)Diam.)
• LS O(N(N-1)/2)) O(N2)
• First iteration, find min cost node from N-1
  nodes not in SPT for 2nd iteration, find min
  cost node from N-2 nodes not in SPT,
• If a sorted list kept, reduce complexity to
  O(Nlog(N))
• After convergence, new routing updates may only
  spur partial recalculations for both DV LS

8
Link State vs. Distance Vector (5)

• Robustness
• Both LS and DV can be completely disabled by a
  single router advertising false/corrupt LSP or DV
• LS can flood false/corrupt LSPs to all routers
• DV can advertise false paths/costs to all
  neighbors
• In ARPANET, malfunctioning routers have
  advertised zero cost, creating a black hole
• In 1997, a bad router in a small ISP advertised a
  false cost, became flooded with traffic,
  disconnecting ISPs from most U.S. backbone
  providers for 3 hours

9
Link State vs. Distance Vector (5)

• Bottom line
• no clear winner in terms of complexity, space,
  robustness,
• but LS is favored in the intra-domain Internet
  due to faster convergence

10
Link-State Cost Metric

• Choice of link cost defines traffic load
• Low cost high probability link belongs to SPT
  and will attract traffic, which increases cost
• Choices for metric
• hop count
• queueing delay
• Transmission delay
• Propagation delay

11
Link-State Cost Metric (2)

• Static metrics (e.g. hop count or another fixed
  cost)
• Less overhead than dynamic metrics
• flood LSPs once initially,
• Thereafter, flood LSPs only if link fails
• Dont have to deal with staleness
• Dynamic metrics can be out of date by the time
  they arrive at a distant router

12
Link-State Cost Metric (3)

• Dynamic metrics take into account
• Changes in link delay (due to congestion)
• Variations in link capacity (wireless BW)
• Dynamic metrics should
• Avoid oscillations
• low cost attracts traffic gt increase cost gt
  less traffic gt low cost gt increase traffic gt
  increase cost
• Achieve good network utilization
• Use link BW efficiently
• Limit overhead from flooding LSPs
• Respond quickly, to avoid performing Dijkstra on
  stale info

13
Original LS ARPANET Metric

• Cost proportional to queue size
• Instantaneous queue length as delay estimator
• Problems
• Did not take into account link speed
• Did not take into account propagation delay of a
  link
• Poor indicator of expected delay due to rapid
  fluctuations
• Moves packets toward smallest queue, rather than
  to destination
• Does not achieve good network utilization

14
New LS ARPANET Metric

• Delay (depart time - arrival time)
  transmission time link propagation delay
• (Depart time - arrival time) captures queuing
• Transmission time captures link capacity
• Link propagation delay captures the physical
  length of the link
• Measurements averaged over 10 seconds
• Update sent if difference gt threshold, or every
  50 seconds
• Achieves better network utilization

15
Problems With New Metric

• Works well for light to moderate load
• Static values dominate
• Oscillates under heavy load
• Queuing dominates
• Congested link advertising high cost pushes
  traffic away gt some links temporarily
  underutilized during heavy load 50 given 2
  links between 2 nodes
• Range is too wide
• 9.6 Kbps highly loaded link can appear 127 times
  costlier than 56 Kbps lightly loaded link
• Can make a 127-hop path look better than 1-hop
• Satellite links penalized, though theyd better
  suit playback video (high BW, non-delay
  sensitive)

16
Revised LS ARPANET Metric

• If a loaded link looks very bad then everyone
  will move off of it
• Want some to stay on to balance load and avoid
  oscillations
• It is still an OK path for some
• Use a hop-normalized metric that
• Has a limited range of values for a given link
• Doesnt jump too much between updates for a given
  link
• Has a limited range of values across different
  link types
• ? gradual change

17
Revised LS ARPANET Metric (2)

• Revised metric is a function of
• Smoothed bounded link utilization
• Link type
• Link utilization
• Measured link util. is sampled over 10sec period
• Link utilization .5current sample .5last
  average
• Normalized according to link type
• Satellite should look good when queuing on other
  links increases

18
Routing Metric vs. Link Utilization
225
9.6 satellite
140
New metric (routing units)
90
9.6 terrestrial
75
56 satellite
60
56 terrestrial
30
0
50
100
25
75
Utilization
19
Observations

• Utilization effects
• High load never increases cost more than 3cost
  of idle link
• Cost f(link utilization) only at moderate to
  high loads
• LSPs flood only if change in Cost exceeds
  threshold
• Link types
• Most expensive link is 7 least expensive link
• High-speed satellite link is more attractive than
  low-speed terrestrial link
• Allows routes to be gradually shed from link

20
Revised LS ARPANET Metric (3)

• Better utilization than earlier two metrics
• Less oscillation
• Allows routes to be gradually shed from link
• But is it sufficiently responsive to avoid stale
  updates?
• Cant propagate updates and react fast enough to
  short time-scale changes in link cost, so
  sluggish response is OK
• Perlman complex.algorithmsbad idea, little
  difference in network capacity between fixed
  cost assignment and dynamic metric

21
Scalability in Internet Routing

• Neither Distance Vector (RIP) nor Link-State
  (OSPF) scale well to many nodes
• DV slow to propagate and converge
• LS overhead of flooding all nodes to all nodes
• DV 50 million nodes gt long distance vectors
• Solution use hierarchy
• Group local routers into a domain, can be a
  subnet on local area or Autonomous System (AS) on
  wide area
• All routers within an AS use an intra-domain
  routing protocol, e.g. RIP and OSPF
• Use another inter-domain routing protocol to
  route between ASs

22
Scalability in Internet Routing (2)
Inter-Domain Routing
AS 1
AS 2
Border/ Gateway Router
Border/ Gateway Router
RIP
OSPF
Intra-Domain Routing
23
Inter-Domain Routing

• Routers within an AS share a common prefix in
  their IP address, i.e. the top 16 bits are all
  the same
• aggregation of routes
• Why Inter-Domain Routing is hard
• 50000 CIDR prefixes to store in each router
• Assigning a cost to the AS between two border
  routers is controlled by owner of each AS
• Advertising a cost of 1000 is fast for one AS1,
  slow for AS2
• Inter-Domain only advertises a reachable path
  across multiple ASs, not shortest path
• Each ASs owner wants per-route QOS/tariffs
• Policy-based routing over performance-based
  routing

24
Border Gateway Protocol (BGP)

• Similar to Distance Vector, but called Path
  Vector instead
• BGP router advertises only reachability info in
  its vector, not costs/hop counts
• E.g. networks 128.96, 192.4.153, and 192.4.3 can
  be reached from AS2
• BGP router advertises its path to each
  destination in its vector
• Avoids loops

25
BGP (2)

• Each routing update carries the entire path
• Loops are detected as follows
• When AS gets route, check if AS already in path
• If yes, reject route
• If no, add self and (possibly) advertise route
  further

R1 to AS 1
R3 to AS 3
R2 To AS 2
26
BGP (3)

• How do intra-domain routers learn about routes to
  other ASs?
• Option 1 Border router injects a default route
  into intra-domain routing protocol for all
  addresses not advertised within AS
• Option 2 Multiple border routers inject a
  specific network prefix into intra-domain routing
  protocol
• E.g. I, BGP 1, have a link to 192.4.54/24 of
  cost X, and I, BGP 2, have a link to
  192.4.72/24 of cost Y

27
BGP (4)

• Perlman Although were probably stuck with BGP
  forever, Ive never been convinced it is the
  right approach.
• Perlman I think the only way to solve the
  general case of policy-based routing is with a
  link state protocol plus source-specified routes.