minimum-cost multicast routing for multi- layered multimedia distribution im phd forum, ntu...
TRANSCRIPT
IM PhD Forum, NTU
Minimum-Cost Multicast Routing for Multi-Minimum-Cost Multicast Routing for Multi-Layered Multimedia DistributionLayered Multimedia Distribution
Hsu-Chen Cheng and Frank Yeong-Sung LinDepartment of Information Management
National Taiwan University
Cheng, Hsu-Chen National Taiwan University 2
Outline Introduction
Multirate Multicasting Modified T-M Heuristic
Two Simple Enhanced Procedures Mathematical Formulation Lagrangean Relaxation
Dual Problem Getting Primal Feasible Solutions
Computational Experiments With Consideration of Link Capacity Constraint Conclusion and Future Research
Cheng, Hsu-Chen National Taiwan University 3
Introduction With the popularity of the Internet, multimedia
multicasting applications such as e-learning and video conference based on network service are growing rapidly.
For Internet service provider, there is one way to achieve the goal of revenue maximization, namely: network planning or traffic engineering.
Traffic engineering is the process of controlling how traffic flows through a network in order to optimize resource utilization and network performance.
Cheng, Hsu-Chen National Taiwan University 4
Introduction (cont.)
S 1N
1D
2N
2D
3D
4D
(a) Example NetworkRate: 2 Mbps
Rate: 0.5 Mbps
Rate: 0.5 Mbps
Rate: 0.5 Mbps
2D
3D
4D
1N 2N
1D
S
Rate: 0.5 Mbps
Rate: 0.5 Mbps
Rate: 0.5 Mbps
Rate: 2 Mbps
(b) Unicast video distribution using point-to-pointtransmission
S 1N
1D
2N
2D
3D
4DRate: 2 Mbps
Rate: 0.5 Mbps
Rate: 0.5 Mbps
Rate: 0.5 Mbps
(c) Multicast video distribution using point-to-multipoint transmission
Cost: 8.5
Cost: 6.5
Cheng, Hsu-Chen National Taiwan University 5
Steiner Tree Problem The minimum cost multicast tree problem, which is the
Steiner tree problem, is known to be NP-complete. The Steiner tree problem is different to the minimum
spanning tree problem in that it permits us to construct, or select, intermediate connection points to reduce the cost of the tree.
Heuristics Shortest Path Tree Based (Minimum Depth Tree, MDT) Spanning Tree Based (MST) T-M Heuristic Others (Optimization based)
Cheng, Hsu-Chen National Taiwan University 6
Steiner Tree Heuristics
MDT
MST
Cost: 8
Cost: 9
S(1) D(2,4)
Cheng, Hsu-Chen National Taiwan University 7
T-M heuristic The T-M heuristic is known to have a solution that is
within a factor of two of the optimum.
Cost: 7
Cheng, Hsu-Chen National Taiwan University 8
Multirate Multicasting Multi-layered Multicasting (Multirate Multicasting)
Taking advantage of recent advances in video encoding and transmission technologies the source only needs to transmit signals that are sufficient for the highest bandwidth destination into a single multicast tree.
A multi-layered encoder encodes video data into more than one video stream, including one base layer stream and several enhancement layer streams.
Cheng, Hsu-Chen National Taiwan University 9
Multirate Multicasting (cont.)
S 1N
1D
2N
2D
3D
4D
(a) Example NetworkRate: 2 Mbps
Rate: 0.5 Mbps
Rate: 0.5 Mbps
Rate: 0.5 Mbps
2D
3D
4D
1N 2N
1D
S
Rate: 0.5 Mbps
Rate: 0.5 Mbps
Rate: 0.5 Mbps
Rate: 2 Mbps
(b) Unicast video distribution using point-to-pointtransmission
S 1N
1D
2N
2D
3D
4DRate: 2 Mbps
Rate: 0.5 Mbps
Rate: 0.5 Mbps
Rate: 0.5 Mbps
(c) Multicast video distribution using point-to-multipoint transmission
B+E
B+E
B
B
B
B
Base layer- 0.5 MbpsEnhancement layer- 1.5 Mbps
Cost: 8.5
Cost: 6.5 Cost: 6
S 1N
1D
2N
2D
3D
4D
Rate: 0.5 Mbps
Rate: 0.5 Mbps
Rate: 0.5 Mbps
(d) Multicast video distribution using multi-layeredMulticasting
Cheng, Hsu-Chen National Taiwan University 10
Multirate Multicasting (cont.) For the conventional Steiner tree problem, the link costs i
n the network are fixed. However, for the minimum cost multi-layered video multicast tree, the link costs are dependent on the set of receivers sharing the link.
The heterogeneity of the networks and destinations makes it difficult to design an efficient and flexible mechanism for servicing all multicast group users.
For (i) a given network topology, (ii) the destinations of a multicast group and (iii) the bandwidth requirement of each destination, we attempt to find a feasible routing solution to minimize the cost of a multicast tree for multi-layered multimedia distribution.
Cheng, Hsu-Chen National Taiwan University 11
Modified T-M Heuristic Takahashi and Matsuyama (1980) proposed an approxi
mate algorithm (T-M Heuristic) to deal with the Steiner Tree Problem.
Maxemchuk (1997) modifies the T-M Heuristic to deal with the min-cost multicast problem in multi-layered video distribution.
Charikar et al. (2004) prove that the problem we discuss here is NP-hard and the result of M-T-M heuristic is no more than 4.214 times of an optimal multicast tree.
Cheng, Hsu-Chen National Taiwan University 12
1 Separates the receivers into subsets according to the receiving rate.
2 Constructs the multicast for the subset with highest rate by using T-M Heuristic.
3 Using the initial tree, the M-T-M heuristic is then applied to the subsets according the receiving rate from high to low until all destination are included in the tree.
Modified T-M Heuristic (cont.)
Cheng, Hsu-Chen National Taiwan University 13
Example I
[0]
[0]
1
2
3 4
1
1
1
1
1[0][0]
Step 5
Rate: 2 Rate: 1
1
2
3 4
1
1
1
1
1
Graph
[1]
1
2
3 4
1
1
1
1
1
[0]
[1]
Step 1
[0]
[0]
1
2
3 4
1
1
1
1
1[1][0]
Step 4Step 3
1
2
3 4
1
1
1
1
1
[0]
[1]
[1][2]
[0]
[1]
1
2
3 4
1
1
1
1
1[1][2]
Step 2
1
2
3 4
Rate: 1
Optimal
Cost: 5 Cost: 4
Cheng, Hsu-Chen National Taiwan University 14
Example II
1
2
3
4
1
3
2
4
3 Rate: 2Rate: 1
Graph
1
2
3
4
1
3
2
4
3
[0]
[3]
[1]
[4]
Step 2
1
2
3
4
1
3
2
4
3
[0]
[3]
[1]
[4]
Step 1
1
2
3
4
1
3
2
4
3
[0]
[3]
[1]
[4]
Step 3
1
2
3
4
1
3
2
4
3
[0]
[3]
[1]
[0]
Step 4
1
2
3
4
1
3
2
4
3
[0]
[3]
[1]
[0]
Step 5
1
2
3
4
1
3
2
4
3
[0]
[0]
[1]
[0]
Step 6
1
2
3
4
1
2 3 Rate: 2Rate: 1
OptimalCost: 11 Cost: 10
Cheng, Hsu-Chen National Taiwan University 15
Two Enhanced Procedures Tie Breaking Procedure
When there is a tie, the node with the largest requirement should be selected as the next node to join the tree.
Drop and Add Procedure 1. Compute the number of hops from the source to the
destinations.
2. Sort the nodes in descending order according to {incoming traffic/its own traffic demand}.
3. In accordance with the order, drop the node and re-add it to the tree. Consider the following possible adding measures and set the best one to be the final tree. Either adds the dropping node to the source node, or to other nodes having the same hop count, or to the nodes having a hop count larger or smaller by one.
Cheng, Hsu-Chen National Taiwan University 16
Example
[0]
[0]
1
2
3 4
1
1
1
1
1[0][0]
Step 5
[1]
1
2
3 4
1
1
1
1
1
[0]
[1]
Step 1
[0]
[0]
1
2
3 4
1
1
1
1
1[1][0]
Step 4Step 3
1
2
3 4
1
1
1
1
1
[0]
[1]
[1][2]
[0]
[1]
1
2
3 4
1
1
1
1
1[1][2]
Step 2
Without TBP
With TBP
[1]
1
2
3 4
1
1
1
1
1
[0]
[1]
Step 1
[0]
[1]
1
2
3 4
1
1
1
1
1[1][2]
Step 2
Rate: 2
1
2
3 4
Rate: 1
Optimal
Rate: 2 Rate: 1
1
2
3 4
1
1
1
1
1
Graph
[0]
[1]
1
2
3 4
1
1
1
1
1[1][2]
Step 3 Step 4
1
2
3 4
1
1
1
1
1
[0]
[1]
[0][0]
Cheng, Hsu-Chen National Taiwan University 17
Mathematical Formulation
Decision Variables Routing Decision: and Traffic Requirement:
Objective Function
Capacity Constraint
min IP l glg G l L
Z a m
gd
gpd gd pl glp P
x m
, ,gg G d D l L (2)
gpdx gly
glm
Cheng, Hsu-Chen National Taiwan University 18
Lagrangean Relaxation
Lagrangian Relaxation Problem
Primal Problem
sub-optimal
subproblem subproblem
sub-optimal
LB < Optimal solution < UB
LBMultiplier
Dual Problem
Adjust uUB
Cheng, Hsu-Chen National Taiwan University 19
Dual Problem We transform the primal problem (IP) into the
Lagrangean Relaxation problem (LR) where Constraints (2) and (6) are relaxed.
LR:
Subproblem 1:
Subproblem 2:
Subproblem 3:
1( , ) min [ ( )]g gd
Sub pl gdl gd gl gpdg G d D p P l L
Z x
2 ( ) min ( )Sub gl g glg G l L
Z D y
3( ) min ( )g
Sub l gdl glg G l L d D
Z a m
( , ) min
g gd g
g gd
D l gl gdl gpd gd pl gdl glg G l L g G d D l L p P g G d D l L
gl gpd pl gl g glg G l L d D p P g G l L
Z a m x m
x D y
Cheng, Hsu-Chen National Taiwan University 20
Getting Primal Feasible Solution
[Lagrangean based modified T-M heuristic]
Step 1 Use as link l’s arc weight and run the M-T-M heuristic.
Step 2 After getting a feasible solution, we apply the drop-and-add procedure described earlier to adjust the result.
gl dgld D
a
Cheng, Hsu-Chen National Taiwan University 21
Computational Experiments Testing Networks:
Regular networks (Grid Network and Cellular Network) Random networks Scale-free networks
For each testing network, several distinct cases which have different pre-determined parameters such as the number of nodes, are considered.
The link cost, destination nodes and traffic demands for each destination are generated randomly.
(a) Grid Network (b) Cellular Network
Cheng, Hsu-Chen National Taiwan University 22
Computational Results
Maximum improvement between M-T-M and LR: Grid networks: 16.18% Cellular networks: 23.23 % Random networks: 10.41% Scale-free networks: 11.02 %
Solution optimality: 60% of the regular and scale-free networks have a gap of less than 10%, but the result of random networks shown a larger gap.
Cheng, Hsu-Chen National Taiwan University 23
Computational Results (cont.) The M-T-M heuristic performed well in many cases, such
as case D of grid network and case D of random network. The tie breaking procedure we proposed is not uniformly
better than random selection. The results of experiments shown that the drop and add
procedure does reduce the cost of the multicast tree.
H 50 1517 1572 1503 1357 1187.332 14.29% 11.79%
D 10 547 547 547 547 541.8165 0.96% 0.00%
Case Dest # M-T-M TB AD UB LB GAP Imp.
Cheng, Hsu-Chen National Taiwan University 24
Computational Results (cont.) There are two main reasons of which the Lagrangean
based heuristic works better than the simple algorithm. First, the simple algorithm routes the group in accordance with
fixed link cost and residual capacity merely, whereas the Lagrangean based heuristic makes use of the related Lagrangean multipliers. The Lagrangean multipliers include the potential cost for routing on each link in the topology.
Second, the Lagrangean based heuristic is iteration-based and is guaranteed to improve the solution quality iteration by iteration.
Cheng, Hsu-Chen National Taiwan University 25
With Consideration of Link Capacity Constraint Single commodity problem Multi-commodity
problem
gd
gpd gd pl glp P
x m
, ,gg G d D l L (1)
gl lg G
m C
l L (2)
Cheng, Hsu-Chen National Taiwan University 26
Multiplier-based Adjustment Procedure (LAP) 1) Compute the aggregate flow of each link.2) Sort the links by the difference between aggregate flow of each link
and the link capacity in descending order.3) Choose the first link. If the difference value of the link is positive, go
to Step 4, otherwise Step 6.4) Choose the group, which have the minimal sensitivity value
on that link, to drop and use as link l’s arc weight and run the M-T-M heuristic to re-add it to the tree. Consider the following possible adding measures and set the best one to be the final tree. Either adds the dropping node to the source node, or to other nodes having the same hop count, or to the nodes having a hop count larger or smaller by one.
5) If a feasible solution is found, go to Step2, otherwise Step 6.6) Stop.
gl gdld D
a
g
l gdld Da
Cheng, Hsu-Chen National Taiwan University 27
Computational Results Maximum improvement between M-T-M and LR-MTM:
Grid networks: 13.46 % Cellular networks: 8.83 % Random networks: 15.4 % Scale-free networks: 10.6 %
Solution optimality: 72% of the regular and scale-free networks have a gap of less than 10%, but the result of random networks shown a larger gap.
Cheng, Hsu-Chen National Taiwan University 28
Conclusion The achievement of this paper can be expressed in term
s of mathematical formulation and experiment performance.
The model can also be easily extended to deal with the constrained multicast routing problem for multi-layered multimedia distribution by adding QoS constraints.
The min-cost model proposed in this paper can be modified as a max-revenue model, with the objective of maximizing total system revenues by totally, or partially, admitting destinations into the system.
Cheng, Hsu-Chen National Taiwan University 29
References Hsu-Chen Cheng and Frank Yeong-Sung Lin, “Minimum-Cost Multicast Routing for
Multi-Layered Multimedia Distribution,” Proc. 7th IFIP/IEEE International Conference on Management of Multimedia and Network Services (MMNS ‘04), Lecture Notes In Computer Science 3271, pp.102-114, October 3- 6, 2004, San Diego, USA.
Hsu-Chen Cheng and Frank Yeong-Sung Lin, “A Capacitated Minimum-Cost Multicast Routing Algorithm for Multirate Multimedia Distribution,” Proc. 2004 International Symposium on Intelligent Signal Processing and Communication Systems (ISPACS ‘04), November 18~19, 2004. Seoul, Korea.
Hsu-Chen Cheng and Frank Yeong-Sung Lin, “Maximum-Revenue Multicast Routing and Partial Admission Control for Multimedia Distribution,” Proc. International Computer Symposium, Dec. 15-17, 2004, Taipei, Taiwan.
Hsu-Chen Cheng and Frank Yeong-Sung Lin, “Maximum-Revenue Multicast Routing and Partial Admission Control for Multirate Multimedia Distribution,” accepted for publication, Proc. The 19th IEEE International Conference on Advanced Information Networking and Applications (IEEE AINA2005).
Frank Yeong-Sung Lin, Hsu-Chen Cheng and Jung-Yao Yeh, “A Minimum Cost Multicast Routing Algorithm with the Consideration of Dynamic User Membership,” accepted for publication, Proc. 2005 The International Conference on Information Networking (ICOIN 2005), Lecture Notes In Computer Science 3288,Jan. 31- Feb 2, Jeju, Korea.
Hsu-Chen Cheng, Chih-Chun Kuo and Frank Yeong-Sung Lin, A Multicast Tree Aggregation Algorithm in Wavelength-routed WDM Networks, Proc. 2004 SPIE Asia-Pacific Optical Communications Conference (APOC ‘04), November 7–11, 2004, Beijing, China.
Q&A
IM PhD Forum, NTU
Minimum-Cost Multicast Routing for Multi-Minimum-Cost Multicast Routing for Multi-Layered Multimedia DistributionLayered Multimedia Distribution