Unicast Routing Tradeoffs

1
Unicast Routing Tradeoffs

• Selma Yilmaz
• Committee Prof. Ibrahim Matta
• Prof. Azer Bestavros
• Prof. John Byers

2
What is Routing?

• Process of finding a path from a source to a
  destination
• Requirements of routing
• Find optimal route
• Scalable
• space complexity topology, resource
  availability, routing table
• time complexity route calculations, processing
  update packets
• communication complexity volume of updates,
  keeping up soft states
• No black hole, loops, oscillations
• Types of routing
• Static, Dynamic
• Inter-domain, Intra-domain
• Hop-by-hop, Source
• Link State, Distance Vector, Path Vector
• Unicast, Multicast

3
Conventional Routing

• Static link metrics, like hop count
• Shortest path routing
• Destination-based only
• Stable, metrics does not change often only in
  case of topology changes
• Connectionless
• No service guarantees
• Plus
• Scales very well
• Minus
• Does not use resources in an efficient way

4
The FISH Problem

• Sub-path R2-R3-R4 may get over-utilized
• Sub-path R2-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 SavageSig99

5
UNICAST ROUTING
ROUTING APPROACH
FORWARDING
ADAPTIVENESS
DOMAIN
DISTANCE VECTOR /PATH VECTOR
LINK STATE
SOURCE
HOP-BY-HOP
STATIC
DYNAMIC
INTRA
INTER
KarInfo00, KodialamInfo01, SuriSV01,




BW CONSTRAINED
QOS ROUTING
ShaikhSig99





BW-DELAY CONSTRAINED
YangICNP01




NeveCC00




MULTIPLE ADDITIVE
MieghemCC01




SalamaInfo97




DELAY CONSTRAINED LEAST COST
GuoMattaICDCS99




YangICNP01, SuriSV01, KarInfo00,
KodialamInfo01
USING INGRESS-EGRESS PAIR INFORMATION




TRAFFIC AWARE
SuriSV01




USING TRAFFIC MATRIX
McQuillanIEEETC80, KhannaSig89




BEST EFFORT




GriffinSig99
6
Outline

• Best Effort
• QoS Routing/Constraint-based Routing
• Traffic Aware Routing
• Ingress-Egress Pair
• Traffic Matrix
• On the Scalability-Performance Tradeoffs in MPLS
  and IP Routing YilmazSpie02
• Future Directions

7
BEST EFFORT
8
Outline

• Best Effort
• Per-packet Dynamic Routing
• Load Balancing Along Equal Length Paths
• Stability
• Inter-domain Routing
• QoS Routing/Constraint-based Routing
• Traffic Aware Routing
• Ingress-Egress Pair
• Traffic Matrix
• On the Scalability-Performance Tradeoffs in MPLS
  and IP Routing YilmazSpie02
• Future Directions

9
Per-packet Dynamic Routing

• Promises
• Avoiding congested links
• Computationally simple
• Distributed
• Stateless
• Scalable
• Difficulties
• Link states change at packet level
• Impractical to generate link state updates at
  packet level
• Larger link state update periods result in
  fluctuations in link state between successive
  updates

10
Per-Packet Dynamic Routing

• ARPANET experience showed that
• choosing right metric is important
• to find optimal paths
• detect congestion, avoid congested areas
• to avoid oscillations
• variations of the link metric
• if goal is to minimize packet delay
• 1stversion instantaneous queue sizeconstant
• delay on links with different characteristics
  looks same
• value changes very rapidly
• poor measure of expected delay on the link
• sub-optimal and unstable paths

11
Per-Packet Dynamic Routing

• McQuillanIEEETC80 average (queuing
  propagationtransmission delay)
• new metric is good predictor of future loads on
  links
• opposite is true at high load
• leads oscillation
• range of delay values are too broad
• makes some links unattractive to all
• no limit on variations reported on successive
  updates for a link
• more update generation
• KhannaSig89 to reduce oscillation, dont
  target only best routes
• use hop normalized metric
• for the same type of links, range of metric value
  is 31
• for all types of links is 71
• static routing at low utilization
• value of a link metric can only change 1/2 hop in
  successive updates
• bounds the the range of oscillations

12
Load Balancing Along Equal Length Paths

• Remedy load balancing inability of static
  routing
• distribute traffic equally along equal cost
  shortest paths
• Ways to achieve this
• Per-packet Round Robin
• not advised PaxsonSig96
• path characteristics may be different
• TCP packets arrive at destination out of order
• unnecessary re-transmissions, waste of bandwidth
• Source destination address based hash
• may not actually balance the load
• Ex OSPF-ECMP
• Static routing if weights are
  administratively assigned
• Ex Cisco suggests 1/link capacity

13
Load Balancing Along Equal Length Paths

• Not always efficient
• not aware of actual loads on the ECMPs
• Need to dynamically assign optimal weights
• But still..
• Conventional IP routing (destination based
  only) is O(N) worse than OSPF-ECMP.
  LorenzDIMACS01

OSPF-ECMP
IP
Per-flow
14
Stability

• Stability Do routes change often? How long does
  it take to reach consistent tables?
• Why it is important?
• routers spend too much time on
• updating their routing tables
• propagating changes
• Static routing is stable
• How to deal with it?
• dont target only the best routes
• Quantizing increases number of target paths
• Use normalized metrics KhannaSig89
• deal with stale link state information
• Update period lt connection arrival and holding
  times
• Use triggered updates
• limit the amount of updates
• per-flow routing instead per-packet

15
Inter-domain Routing

• What if source and destination are in different
  ASs?
• based on min-hop AS path routing
• there is no universally agreed metric among ASs
• each AS may have its own criteria for path
  selection
• BGP is the current inter-domain routing protocol
• path vector, policy based
• BGP allows each AS
• independently formulate its routing policies
• overwrite distance metrics in favor of policy
  constraints
• BGP is not pure distance vector
• may diverge, result in persistent oscillation in
  the global Internet
• not robust to failures

16
Inter-domain Routing

• statically analyzing convergence of BGP (even for
  known policies and single destination) is
  impractical GriffinSig99
• Inter-domain QoS routing is more difficult than
  intra-domain
• no universally agreed metric among Ass
• policies overwrites metrics
• scale of topology

17
QOS ROUTING/CONSTRAINED-BASED ROUTING
18
Outline

• Best Effort
• QoS Routing/Constraint-based Routing
• What is QoS Routing?
• Objectives of QoS Routing
• Difficulties with QoS Routing
• Ways to Deal with Increased Cost of QoS Routing
• Closer Look at Some QoS Performance-Cost
  Tradeoffs
• Effects of Topology
• Effects of Traffic Characteristics
• QoS Routing Problems
• A Possible Classification of QoS Routing Problems
• Multiple Constraints Routing Problem
• Delay Constrained Least Cost Routing
• Can we achieve QoS guarantees at lower cost?
• Improving Long Term Utilization of Network
• Traffic Aware Routing
• Ingress-Egress Pair
• Traffic Matrix
• On the Scalability-Performance Tradeoffs in MPLS
  and IP Routing YilmazSpie02

19
What is QoS Routing?

• QoS Routing Problem Definition
• Given a network graph G(V,E)
• a source s and a destination d
• a set of QoS constraints C
• possibly an optimization goal
• Find the best feasible path from s to d, which
  satisfies C
• Example
• Delay at most 4 gt s -gt k -gtl -gtt
• Bandwidth at least
  1.5 gt s -gt i -gt l -gt d
• gt s -gt k -gt l -gt d
• Both gt s -gt k -gtl -gtt
• Components 1. Maintain topology and link state
  information
• 2. Distribution of link state
  information
• 3. Route computation

(bandwidth,delay)
j
20
Objectives of QoS Routing

• Provide guarantees for end-to-end performance
• Can identify feasible paths better
• dynamic routing, accurate information about
  available resources
• aware of QoS requirements of the requests
• Resource reservation
• Perform admission control
• Improve the long term utilization of the network
• Perform admission control
• Prefer social paths
• Use network resources in such a way that
  probability of accepting future requests is
  increased
• Use the network resources in an efficient way
• Prefer shortest paths to limit resource
  consumption
• Prefer least loaded path to load balance

21
Difficulties with QoS Routing

• Performance depends on accuracy of the network
  state information
• changes frequently
• flooding costs
• tradeoff between communication overhead and
  accuracy
• More sophisticated route computations
• to satisfy multiple constraints (some
  combinations are NP-hard)
• tradeoff between simpler path computations and
  better quality paths
• More frequent route computations
• maybe on-demand
• tradeoff between per-request computational cost
  and quality of path
• Paths need to be pinned and maintained afterwards
• per-flow state
• signaling cost to set-up, tear down, keep alive,
• routing table entry larger tables, slower lookups

22
Ways to Deal with Increased Cost of QoS Routing

• Reduce volume of updates by controlling
• when/how often send
• periodic updates, threshold/class-based triggers,
  clamp-down timers
• what to send
• only the specific link, or all the nodes links,
• absolute or quantized value
• Use heuristics for simpler route computations
• Use pre-computation or path caching instead of
  on-demand
• Increase granularity, like class-based
• Use hierarchical routing
• Use hybrid routing Restrict QoS routing only a
  group of flows and use static routing for the rest

23
Closer Look at Some QoS Performance-Cost
Tradeoffs

• 1. How Often Paths are Computed?
• On-demand, pre-computation, path caching
• 2. Type of Computation
• Hop-by-hop, source, crank-back
• 3. Granularity
• Per-flow, larger aggregates
• 4. Absolute vs. quantized values
• 5. Hierachical vs Flat
• 6. Hybrid Routing StaticDynamic
• 7. Link state, Distance Vector, Path Vector

24
Closer Look at Some QoS Performance-Cost
Tradeoffs

• 1. How Often Paths are Computed?
• On-demand per-request
• plus
• Exact QoS requirements and destination are known
  at path selection
• Yields better routes using the most recent link
  metrics available
• Reduces space complexity
• minus
• Increases computational complexity
• Per-request processing overhead is high
• If large clamp down timer is used, re-discovers
  same paths
• use pre-computation

25
Closer Look at Some QoS Performance-Cost
Tradeoffs

• Pre-computation
• periodic/after receiving certain number of
  updates
• plus
• Reduces per-request processing overhead
• minus
• Must compute all possible paths to all
  destinations for all possible QoS requirements
• Must store pre-computed paths
• Adds path selection cost Given destination and
  bandwidth find a suitable path from QoS table
• Some of the paths may never be used
• Periodic pre-computation results in routing
  performance loss ApostolopoulosSig98
• For sensitive triggers maybe more costly than
  on-demand

26
Closer Look at Some QoS Performance-Cost
Tradeoffs

• Hybrid Path Caching ApostolopoulosICNP98
• Remember k previously computed paths to a
  destination
• Choose shortest feasible
• plus
• good balance between per-request processing
  overhead and quality of paths
• ability of load balance/pack
• the more entries, the better results
• bigger cache size
• update paths instead of invalidate
• Small cache size for faster lookup and less
  storage
• minus
• cache need to maintained entries must be
  updated, invalidated etc.
• for small update periods, cost can exceed on
  demand
• storage overhead at each node per destination

27
Closer Look at Some QoS Performance-Cost
Tradeoffs

• 1. How Often Paths are Computed?
• On-demand, pre-computation, path caching
• 2. Type of Computation
• Hop-by-hop, source, crank-back
• 3. Granularity
• Per-flow, larger aggregates
• 4. Absolute vs. quantized values
• 5. Hierachical vs Flat
• 6. Hybrid Routing StaticDynamic
• 7. Link state, Distance Vector, Path Vector

28
Closer Look at Some QoS Performance-Cost
Tradeoffs

• 2. Type of Computation
• Hop-by-hop
• path computation is distributed
• scalable
• occasionally suffer from loops due to
  inconsistencies in the path discovered by
  different routers
• speeds up the establishment of a path by avoiding
  the route computation at the source
• generally used for pre-computation

29
Closer Look at Some QoS Performance-Cost
Tradeoffs

• Source routing
• plus
• guarantees loop free routing
• enforces the route computed at the source
• allows sophisticated and complex path selection
  algorithms with diverse resource requirements
• generally used for on-demand
• minus
• has scalability problem
• centralized computation
• global state needs to maintained at each node
• the path is included in the header of the packet
• each router has to process the packet

30
Closer Look at Some QoS Performance-Cost
Tradeoffs

• Hybrid Crank-back (PNNI)
• route using source routing
• in case of failure during path establishment
  phase, use a local search at the point of failure
  
• a way to handle inaccurate network state
  information
• increases time to set up a path for an incoming
  flow

31
Closer Look at Some QoS Performance-Cost
Tradeoffs

• 1. How Often Paths are Computed?
• On-demand, pre-computation, path caching
• 2. Type of Computation
• Hop-by-hop, source, crank-back
• 3. Granularity
• Per-flow, larger aggregates
• 4. Absolute vs. quantized values
• 5. Hierachical vs Flat
• 6. Hybrid Routing StaticDynamic
• 7. Link state, Distance Vector, Path Vector

32
Closer Look at Some QoS Performance-Cost
Tradeoffs

• 3. Granularity
• Per-flow
• plus
• better load balance, better long term performance
  
• better service guarantees
• minus
• expensive
• path computation occurs more frequently
• network state changes more frequently
• requires smaller update trigger
• state must be maintained for each flow
• signaling cost path set-up, tear down, keep
  alive
• routing table entry size, and lookup

33
Closer Look at Some QoS Performance-Cost
Tradeoffs

• Flows can be larger aggregates of multiple flows,
  like class-based, or destination based
• plus
• cheaper less state, smaller routing tables,
  faster lookup
• more scalable
• minus
• demands with larger bandwidth requirements have
  higher blocking probability MaICNP97 ,
  MattaInfo98 , ShaikhICNP08
• bandwidth fragmentation
• may increase potential of instability
• large volumes of traffic will be shifted from one
  place to the other
• activates update triggers more often

34
Closer Look at Some QoS Performance-Cost
Tradeoffs

• 1. How Often Paths are Computed?
• On-demand, pre-computation, path caching
• 2. Type of Computation
• Hop-by-hop, source, crank-back
• 3. Granularity
• Per-flow, larger aggregates
• 4. Absolute vs. quantized values
• 5. Hierachical vs Flat
• 6. Hybrid Routing StaticDynamic
• 7. Link state, Distance Vector, Path Vector

35
Closer Look at Some QoS Performance-Cost
Tradeoffs

• 4. Absolute vs. quantized values
• Quantized
• may reduce accuracy especially if only a few
  quantized values are being used
• increases number of equal cost paths
• by not targeting only a single best choice,
  increases stability
• help the path selection being stuck with a single
  bad choice
• influences when the next update will take place
• allows smaller routing table size

36
Closer Look at Some QoS Performance-Cost
Tradeoffs

• 1. How Often Paths are Computed?
• On-demand, pre-computation, path caching
• 2. Type of Computation
• Hop-by-hop, source, crank-back
• 3. Granularity
• Per-flow, larger aggregates
• 4. Absolute vs. quantized values
• 5. Hierachical vs Flat
• 6. Hybrid Routing StaticDynamic
• 7. Link state, Distance Vector, Path Vector

37
Closer Look at Some QoS Performance-Cost Tradeoffs

• 5. Hierachical vs Flat
• Flat
• nodes have full knowledge of topology
• Does not scale
• cost of communication, computation, and storing
  overhead is huge
• Hierachical PNNI, OSPF (only two level)
• reduce the size of network by aggregating
  topology
• Each node knows
• complete topology within its group
  
• summarized information of parent
  group
• Link
  aggregation
• A.1-A.2 is aggregate of
• 
  A.1.3-A.2.3 and A.1.1-A.2.2

Parent Group
Border routers
38
Closer Look at Some QoS Performance-Cost Tradeoffs

• 5. Hierachical vs Flat
• plus
• can scale to large networks
• amount of information that is stored is reduced
• amount of information that is distributed is
  reduced
• minus
• tradeoff between amount of aggregation and
  accuracy
• loss of detailed information
• state of logical links are combination of many
  lower-level links
• as states are aggregated, imprecision is also
  aggregated
• effects quality of selected paths

39
Closer Look at Some QoS Performance-Cost
Tradeoffs

• 1. How Often Paths are Computed?
• On-demand, pre-computation, path caching
• 2. Type of Computation
• Hop-by-hop, source, crank-back
• 3. Granularity
• Per-flow, larger aggregates
• 4. Absolute vs. quantized values
• 5. Hierachical vs Flat
• 6. Hybrid Routing StaticDynamic
• 7. Link state, Distance Vector, Path Vector

40
Closer Look at Some QoS Performance-Cost Tradeoffs

• 6. Hybrid Routing StaticDynamic
• Classify flows according to their
  characteristics and dynamically
• /statically route some classes
• Use dynamic per-flow routing only for
  long-lived flows
• Route class of short-lived flows statically
  ShaikhSig99
• plus
• increases stability
• per-flow is more stable than per-packet
• class-based is more stable than per-flow
• link states change more slowly
• number of flows that are dynamically routed is
  decreased
• flow duration is increased
• frequency of updates are reduced

41
Closer Look at Some QoS Performance-Cost Tradeoffs

• 6. Hybrid Routing StaticDynamic
• slower link state change and less flows to be
  dynamically routed reduces overheads
• number of route computation
• number of per-flow state
• number of signaling opertions pinning, keeping
  up soft state, tearing down
• smaller forwarding tables
• more robust to stale link state information
• flow trigger value can be used to
• balance cost and performance tradeoff
• balance stability and adaptiveness
• minus
• under accurate link state information,
  performance is worse than dynamic routing
• short flows have to be on static paths

42
Closer Look at Some QoS Performance-Cost
Tradeoffs

• 1. How Often Paths are Computed?
• On-demand, pre-computation, path caching
• 2. Type of Computation
• Hop-by-hop, source, crank-back
• 3. Granularity
• Per-flow, larger aggregates
• 4. Absolute vs. quantized values
• 5. Hierachical vs Flat
• 6. Hybrid Routing StaticDynamic
• 7. Link state, Distance Vector, Path Vector

43
Closer Look at Some QoS Performance-Cost Tradeoffs

• 7. Link state, Distance Vector, Path Vector
• Link state OSPF, IS_IS
• each router knows the entire network topology
• plus
• allows more complex path selection algorithms
• minus
• must maintain topology database at each node
• flood updates
• does not scale well to large networks
• use hierarchical topology aggregation
• Distance Vector RIP
• each node keeps shortest path tree rooted at
  itself to destinations
• plus
• requires less memory
• more scalable

44
Closer Look at Some QoS Performance-Cost Tradeoffs

• minus
• cannot support sophisticated path computations
• may have convergence problems
• does not allow computation routes specific to the
  requirements of an individual flow
• Path Vector BGP
• routing tables not only keeps
  ltdestination,nexthop,costgt but also the
• corresponding path
• plus
• eliminates loops
• minus
• needs more storage to keep paths
• the paths needs to be processed

45
Effects of Topology on QoS

• Overheads of computing routes and distributing
  link states
• Specifies number of candidate paths between each
  pair
• better chances of load balance
• Larger hop count connections are harder to route
  under inaccurate link state information
  ShaikhICNP98
• Pruning should be disabled for loosely connected
  topologies ShaikhICNP98 ApostolopoulosSig98
• Specifies relative cost of single destination and
  all-destination path computation

46
Effects of Traffic Characteristics on QoS

• Connection inter-arrival time and holding time
• Frequency of link state updates
• Large update periods relative to arrival rates
  and hold time leads to flapping
• Longer-lived flows allows use of larger link
  state update period
• ShaikhSig99
• Exponential holding times gives lower blocking
  probability than Pareto with the same mean for
  the same link state update periods
• With increasing number of short-lived flows,
  inaccuracy increases
• MaICNP97 ,ShaikhICNP98

47
Effects of Traffic Characteristics on QoS

• Requested Bandwidth
• High bandwidth flows have higher blocking
  probability
• Low bandwidth flows can more easily use the
  available bandwidth
• more robust to inaccuracies
• High bandwidth flows cause large fluctuations in
  link state
• needs more frequent update messages

48
Effects of Traffic Characteristics on QoS

• Uniform vs Non-uniform Traffic
• Under non-uniform load, QoSR performs better than
  static
• Under uniform traffic, QoSR may perform worse
  than static
• depending on the update period staleness causes
  route flapping
• ShaikhICNP98
• very accurate link state information causes
  excessive alternate routing
• uses longer paths with extra resources
• interfere with min hop traffic computing for the
  same links
• MattaInfo98, ApostolopoulosSig98
• Solution Trunk reservation

49
QoS Routing Problems

• Link metrics Additive delay, cost
• Multiplicative probability
  of successful transmission
• Concave bandwidth, buffer
  space
• Path metrics
• Additive Sum of link metrics along the
  path
• Concave Max(Min) of link metrics along
  the path
• Find a path that optimizes path metric
• for concave metric Link optimization
• for additive/multiplicative metric Path
  optimization
• Find a path whose path metric is above/below a
  specified value
• for concave metric Link constrained
• for additive metric Path constrained
• ChenIEEEN98

50
A Possible Classification of QoS Routing Problems

• basic routing problems
  composite routing problems
• link optimization link-constrained
  link-optimization
• 
  
• link-constrained path-optimization
• link-constrained
• multiple link-constrained
• link-constrained path-constrained
• path-constrained link-optimization
• path-optimization
• path-constrained path-optimization
• NP-complete
• path-constrained multi-path-constrai
  ned
• NP-complete

51
Multiple Constraints Routing Problem

• Problem Definition
• Given
• a network graph G(V,E)
• vector of m additive link metrics for each edge
• a source node s
• a destination node d
• vector of m positive constraints, L
• and
• Path P(s,k,..,l,d) is specified by a vector
• Find a path satisfying constraints such that
  li(P)ltLi for all i1,2,..,m
• How to run shortest path algorithm?

52
Multiple Constraints Routing Problem

• Path Length assuming m2
• Linear l1(P)a l2(P)ß
• Non-linear max ( l1(P)/L1, l2(P)/L2)
  NeveCC00 MieghemCC01
• Linear Representation of Path Length
• 
  Problems
• 1. May fail
  
  
• 
  Optimal path according to the
• 
  new weight function maybe infeasible
• 
  2. How to choose weights?
• 
  

Solution returned by Dijkstra
53
Multiple Constraints Routing Problem

• Non-linear Representation of Path
• Length
• Problem
• subsections of shortest paths are not
  necessarily shortest paths
• Dijkstra fails to find shortest paths

54
Multiple Constraints Routing Problem

• Example
• From source to u, Dijkstra will choose P2
  since the path length of P2
• Length of P2max(5/12,5/12)5/12 lt
  Length of P1max(10/12,1/12)10/12
• Shortest path from source to destination is
  through P1, since
• max(11/12,11/12)ltmax(6/12,15/12)
• Solution
• Use k shortest path algorithm
• At each step of Dijkstra, store k shortest paths
• kmaxminL1L2..Lm/max(Li), floor(e(N-2)!)
• tunable tradeoff between accuracy and complexity

55
Delay Constrained Least Cost Routing
LDP
Path Cost

• Problem Definition
• Given
• a network graph G(V,E)
• nonnegative cost C(e) for each edge e
• nonnegative delay D(e) for each edge e
• a source node s
• a destination node d
• positive delay constraint ?d
• Find a path satisfying
• min Cost(Pi) and Pi ?P(s,d)
  iff Delay(Pi)lt ?d
• Pi ?P(s,d)
• where for a path P vo,v1,..,vn cost of a
  path P Cost(P) Se in P C(e)
• delay of a path P Delay(P)
  Se in P D(e)

Optimal Path
LCP
?d
Path Delay
56
Delay Constrained Least Cost Routing

• Example ?d 3, delay/cost

destination
source
Optimal DCLC path delay3,cost5
Least Delay Path delay2,cost6
Least Cost Path delay4,cost4 infeasible
57
Delay Constrained Least Cost Routing

• Convert DCLC Routing Problem to DCC
• GuoMattaICDCS99,COMNET02
• What will be the cost bound?
• Path Length
• Path Delay/(1- Path Cost/Cost Bound)) if
  feasible
• infinity if not feasible
• which gives preference to low cost paths
• Example
• ?cCost(LDP)6 LDP is AGE
• Look at the paths with smaller cost
• AGE Cost6, Delay2
• ABCE Cost6, Delay3
• ABCDE Cost4, Delay4, Path Lengthinfinity
• AGFE Cost5, Delay3, Path Length18

58
Issues

• Non-linear path length requires k-shortest path
  algorithm
• Added space complexity is O(kV)
• With non-integer constraints, k is big O(N!)
• run time complexity, O(kVlogkVk2mE), is no
  longer polynomial
• if choose smaller value for k, algorithm is not
  exact
• may miss the shortest path
• Reduced search space GuoMattaICDCS99,COMNET02
• eliminate infeasible links by assigning infinite
  weight
• find tighter cost bound, run BlokhGutin
• Better chance of finding optimal
• smaller k may achieve good enough
• run time complexity increased because of
  pre-processing step
• Hop-by-hop computation of DCLC

59
Delay Constrained Least Cost Routing

• Distributed solution
  ?d 3, d/c
• Each node keeps Cost and Delay Vectors
• Destination/least cost(delay) from source
  to destination/nexthop
• At each hop, choose either LCP/LDP
• Active node chooses LCP if
• delaysofardelayD(LDP)lt ?d
• otherwise chooses LDP
• Better scalability, and time complexity
• Increased communication complexity
• Still per-flow state

D
D
S
60
Can we achieve QoS guarantees at lower cost?

• QoS routing generally
• asks for source routing, link state, on-demand
  mode,
• requires per-flow state
• NOT SCALABLE!
• Is it possible to achieve sufficient level of QoS
  guarantees with hop-by-
• hop destination based only routing
  (connectionless) ?
• HbHDBO QoS cannot guarantee to find exact results
  where
• routes are computed from intermediate node to
  destinations
• only next hop information is kept
• MieghemCC01
• Simulations shows that quality of HBHDBO QoS
  routing is good
• 90 of the time exact solution is found
• With active networking HBHDBO QoS can be done
• each packet carries remaining constraints
• each packet carries path traversed so far , and
  constraints

61
2 5 5
5 4 6
d
4 1 7
b
3 7 1
2 1 2
f

3 3 2
a
e
i
5 3 8
2 2 7
2 3 9
(2352)/140.86 (1333)/110.91 (2289)/220
.95 i chooses f as next hop
3 5 4
g
c
7 8 2
2 5 5
5 4 6
2 5 5
5 4 6
(43)/140.5 (17)/110.73 (71)/220.36 e
chooses b as next hop
d
Hop-by-hop path (2333)/140.86 (1371)/111.0
9 (2217)/220.55
4 1 7
d
4 1 7
b
b
3 7 1
2 1 2
3 7 1
2 1 2
f
3 3 2
f
3 3 2
a
e
a
e
i
5 3 8
i
2 2 7
5 3 8
2 2 7
Shortest Path (2352)/140.86 (1333)/110.91
(2289)/220.95
(52)/140.5 (33)/110.54 (98)/220.77
2 3 9
3 5 4
2 3 9
3 5 4
g
c
7 8 2
g
c
7 8 2
62

TAMCRA NeveCC00
SAMCRA MieghemCC01
DCUR SalamaInfo97
SSRDCCR GuoICDCS99
Solves
Generic Multiple Additive
Generic Multiple Additive
Delay Constrained Least Cost
Delay Constrained Least Cost
Source
Hop-by-hop
Routing Strategy
Source
Hop-by-hop
Type
Link State or Distance Vector
Link State or Distance Vector
Link State
Link State
O(hVlog(hV)h2mE)) per-flow h is accuracy tuning
index m is number of metrics
O(hVlog(hV)h2mE)) per-packet h is accuracy
tuning index m is number of metrics
O(hVlog(hV)h2mE)) per-flow BG O(cElogV) c
of iterations h is accuracy tuning index m is
number of metrics

O(1)
Time Complexity
O(k) to distribute link states if Link State k is
number of neighbors O(V) if Distance Vector O(V3)
O(k) to distribute link states k is number of
neighbors
O(k) to distribute link states if Link State k is
number of neighbors O(V) if Distance Vector
O(k) to distribute link states k is number of
neighbors
Communication Complexity
O(E) for network state O(hmN) for h shortest
path h is accuracy tuning index
O(E) for network state if Link State O(kV) for
path state if Distance Vector k is number of
neighbors O(hmN) for h shortest path m is number
of metrics
O(E) for network state if Link State O(kV) for
path state if Distance Vector k is number of
neighbors O(V) for cost vector O(V) for delay
vector
O(E) for network state O(hmN) for h
shortest path h is accuracy tuning index
Space Complexity
Per-flow
none
Per-flow
Per-flow
Maintained State
none
none
Extra Information Used
none
none
63
Improving Long Term Utilization of Network

• Choose paths that uses network resources in an
  efficient way
• Increase revenue, decrease blocking
  probability
• Use additional constraints
• limit resource consumption shortest paths
• improves performance under heavy load
  MaICNP97
• load balance use alternate (wider) paths
• improves performance under light load MaICNP97
• increases blocking probability of high bandwidth
  requests
• MattaBestavrosInfo98
• good for evenly distributed load
• load packing most utilized paths (best fit)
• approximates perfect fit, which is NP-hard
• minimize fragmentation
• improves fairness for large bandwidth requests
• bad for uniform traffic MattaBestavrosInfo98
• extremely sensitive to link state inaccuracies

64
Improving Long Term Utilization of Network

• load profiling MattaBestavrosInfo98 try to
  match load profile and bandwidth availability
  profile
• more robust to routing inaccuracies, allows more
  longer update periods

65
TRAFFIC AWARE ROUTING
66
Outline

• Best Effort
• QoS Routing/Constraint-based Routing
• Traffic Aware Routing
• Ingress-Egress Pair
• Traffic Matrix
• On the Scalability-Performance Tradeoffs in MPLS
  and IP Routing YilmazSpie02
• Future Directions

67
What is Traffic Aware Routing?

• Input available resources and some knowledge
  about traffic like ingress-egress pairs, or
  traffic matrix
• Goal optimize/maximize network usage or service
  guarantees
• Ingress-egress Pairs location of
  source-destination traffic
• allows efficient use of network resources,
  maximize revenue
• Traffic Matrix demands between
  source-destination pairs
• allows to pre-compute set of routes optimizing
  some performance metric
• based on long-term averages, e.g. daily
• need to be re-optimized
• static for a period of time
• generally done off-line in a centralized manner
• not designed to handle traffic fluctuations in
  real-time
• Routing tables converge slowly

68
Is achieving good long-term performance enough?

• Traffic matrix is for long term traffic
  characteristics
• Used to optimize long term performance
• What about short term performance?
• Short-term performance may be sub-optimal
  SridharanITC01
• Traffic and link loads over shorter time scales
  fluctuate around the long term average values
• finer granularity, shorter time scales gt more
  variability
• Compute a new set of optimal routes at shorter
  time scales to avoid overloading
• not feasible, not stable
• Use traffic aggregates
• coarser granularity restricts load balancing
  ability
• poorer long term performance

69
Ingress-egress Pairs
70
How to use this extra information to improve
performance?

• Smarter usage of resources for QoSR
• increase utilization and long term performance
• on-demand, no knowledge of future requests, no
  splitting of flows
• Bandwidth Constrained
• Among feasibles, pick paths that interfere
  least with future requests
• KarInfo00
• How to find minimum interference paths?
• Maxflow Upperbound on the total amount
  of bandwidth that can be
• routed between an ingress-egress pair

71
How to use this extra information to improve
performance?

• Can we use LP to maximize the minimum/sum of
  maxflow/s between other ingress-egress pairs?
• Unsplitability restriction makes the problem
  NP-hard
• Heuristic MIRAKarInfo00
• Assign weights that are increasing function of
  criticality
• Routing over a critical link
  decreases the maxflow value of some
• ingress-egress pairs
• defer loading of critical links whenever possible
  
• By Dijkstra, find shortest path
• Maxflow23
• Mincut(s,v1,v2,v4,v3,d)
• Capacity of mincut23
• Critical links (v1,v3), (v4,v3), (v4,d)
• 
  

72
Traffic Matrix
73
How to use traffic matrix to improve performance?

• Optimal Solution solve LP formulation to
  optimize a metric like minimizing maximum link
  utilization
• allows unrestricted split
• can only be simulated by directing packets along
  logical connections
• costly, more than per-flow state is needed in
  case of split
• not scalable
• Calculate OSPF Weights
• Assign weights in such a way that shortest
  path routing will prefer the underutilized links
• plus
• no per-flow state
• simple shortest path routing
• scalable

74
How to use traffic matrix to improve performance?

• Unrestricted Case WangInfo01
• 1. solve LP formulation to optimize a metric
• 2. solve dual of LP to find link weights that
  with shortest path computation will reproduce
  optimal routes obtained in (1)
• Minus current OSPF only supports equal split
• OSPF-ECMP
• cannot be formulated as an LP
• equal split along equal cost paths makes the
  problem NP-hard FortzInfo00
• find heuristics to set weights
• local search heuristic FortzInfo00
• set weights as an exponential function of link
  load LorenzDIMACS01

75
How to use traffic matrix to improve performance?

• Use both traffic matrix and ingress-egress pair
  information
• Offline (pre-processing) phase
• Compute an optimal solution by using LP for
  traffic matrix
• solve multi-commodity flow problem
• no restriction on splitting
• each profile is a commodity
• result is pre-allocated capacities for each
  commodity
• Online phase Use the result of offline phase as
  virtual capacities to route individual requests
• Suri01

76
Can clever weight setting for OSPF-ECMP replace
per-flow routing?

• Weight setting for OSPF-ECMP cannot replace
  MPLS as a TE tool.
• OSPF-ECMP is O(N) worse than per-flow routing
  with respect to
• maximum throughput and maximum utilization that
  can be achieved.
• LorenzDIMACS01 FortzInfo00
• What about the gap in practice?
• may be not too bad for realistic topologies and
  traffic patterns
• worst-case may never be observed in practice
• what aspects of topology makes gap less
  significant
• what aspects of traffic patterns make gap less
  significant?

77
Demands are 1unit
If only one link is used at all nodes, then
maximum throughput1
If 2 links are used at some nodes, then
maximum throughput2
OSPF-ECMP
Per-flow maximum throughputN
78
...
Sources
1
1
1
u1
CN-1
CN-2
u2
u3
C1
CN-3
C1
N
...
C1
vNuN
v2
v1
v3
N
Destination
d
If only one link is used at all nodes, then
maximum utilizationN
If 2 links are used at some nodes, then maximum
utilizationN/2
OSPF-ECMP
Per-flow maximum utilization1
79
MDWCRAYangICNP01
MOCAKodialamInfo01


Solves
Bandwidth-Delay Constrained
Guaranteed Bandwith


Source
Source
Source
Hop-by-hop
Routing Strategy
Type
Link State
Link State
Link State
Link State or Distance Vector

O(nVE2) n is number of ingress-egress pairs
O(VE) for on-line phase offline phase
O(pE2logVp Vlog(pV)) p number of
ingress-egress pairs
O(1)
Time Complexity
O(k) to distribute link states k is number of
neighbors
O(k) to distribute link states k is number of
neighbors
O(k) to distribute link states if Link State k is
number of neighbors O(V) if Distance Vector
O(k) to distribute link states k is number of
neighbors
Communication Complexity
O(E) for network state O(V2) for ingress-egress
pair matrix
O(E) for network state O(V2) for ingress-egress
pair matrix O(V2) for traffic profile matrix
O(E) for network state if Link State O(kV) for
path state if Distance Vector k is number of
neighbors
O(E) for network state O(pV) to keep paths with
different weights
Space Complexity
Per-flow
Per-flow
Per-class and per-flow
none
Maintained State
Ingress-egress pair matrix Traffic profile matrix
none
Extra Information Used
Ingress-egress pair matrix
Ingress-egress pair matrix
80
Outline

• Best Effort
• QoS Routing/Constraint-based Routing
• Traffic Aware Routing
• Ingress-Egress Pair
• Traffic Matrix
• On the Scalability-Performance Tradeoffs in MPLS
  and IP Routing YilmazSpie02
• Future Directions

81
On the Scalability-Performance Tradeoffs in MPLS
and IP Routing YilmazSpie02

• How far are MIRA and PBR from Optimal?
• Does increased complexity mean better
  performance? Scalability?
• Compare the performance of
• Optimal, WSP, MIRA,
  PBR
• Performance metrics
• bandwidth acceptance ratio
• utility
• maximum utilization

82
On the Scalability-Performance Tradeoffs in MPLS
and IP Routing YilmazSpie02

• Optimal per-packet routing Optimizes for
  bandwidth acceptance ratio
• Solve multicommodity flow problem at each flow
  arrival/departure where
• each flow is a commodity
• xi(e) amount of commodity i routed through
  edge e with appropriate constraints
• Excess edges (with costinfinity) are added
  to always have feasible solution

83
On the Scalability-Performance Tradeoffs in MPLS
and IP Routing YilmazSpie02

• An individual flow can be split, get different
  bandwidth values, assigned to different paths
  during its lifetime
• Prefers shorter paths (less costly)
• Does not load balance

Possible bandwidth assignment for a flow asking
for b units of bandwidth. Accepted amountb-b4
84
On the Scalability-Performance Tradeoffs in MPLS
and IP Routing YilmazSpie02

• Plus
• Avoid congested links, better utilization and
  performance
• Computationally simple
• Does not use extra information like
  ingress-egress pairs or traffic matrix
• Distributed
• Stateless
• Scalable
• Minus
• Cannot provide guaranteed service
• Hard to achieve in practice
• routing oscillation

85
On the Scalability-Performance Tradeoffs in MPLS
and IP Routing YilmazSpie02

• Widest-Shortest Path
• Choose feasible min-hop path
• Break ties by picking the widest
• Improves performance by
• limiting resource consumption
• balancing load
• MaICNP97

86
On the Scalability-Performance Tradeoffs in MPLS
and IP Routing YilmazSpie02

• MIRA KarInfo00
• To route a demand (a,b,D)
• 1. For all ingress-egress pairs (s,d) ! (a,b)
• Compute maxflow values
• Compute critical links that belong to all
  possible mincuts
• Compute weights w(l)S(s,d)l is critical asd
  
• 2. Eliminate all links with residual bandwidth lt
  D
• 3. Use Dijkstra to compute shortest path from a
  to b
• 4. Route the demand and update residual
  capacities
• Some choices for asd
• Can be chosen to reflect the importance of pairs
• If asd1, w(l) shows number of ingress-egress
  pairs for which link is critical
• If asd1/ maxflow value for ingress-egress pair
  (s,d), then critical links for ingress-egress
  pairs whose maxflow value is lower will weigh
  heavier

87
On the Scalability-Performance Tradeoffs in MPLS
and IP Routing YilmazSpie02

• Problems with MIRA
• Computationally expensive
• Time ComplexityO(pVE2) for each request
• pnumber of ingress-egress pairs
• Ingress-egress pair matrix need to be maintained
• Per-flow state
• No admission control
• Should be able to reject a request even if
  there is a feasible path, if accepting will
  result in high blocking probability for future
• demands

88
On the Scalability-Performance Tradeoffs in MPLS
and IP Routing YilmazSpie02

• Profile-Based Routing Suri01
• Traffic Profile (classID, si, si, Bi) Aggregate
  expected traffic between
• 
  ingress si -egress di for a class classID.
• Each class is separate commodity.
• Offline phase
• find optimal distribution of profiles
• use result as virtual capacity for each class
• Online phase
• route individual requests using pre-allocated
  capacities
• Offline phase serves as admission control
  mechanism
• may reject a request even if there is a feasible
  path if a