Optical Interconnection Networks for Scalable High-performance Parallel Computing Systems
Scalable Traffic Grooming in Optical...
Transcript of Scalable Traffic Grooming in Optical...
Scalable Traffic Grooming in Optical Networks
George N. Rouskas
Department of Computer Science
North Carolina State University
Joint work with: Zeyu Liu, Hui Wang
Funded by the NSF under grant CNS-1113191
ACP 2012 – November 10, 2012 – p.1
Outline
Motivation and Challenges
Scalable Optical Network Design
Routing and Wavelength assignment (RWA)
Traffic Grooming
Conclusion and Future Directions
ACP 2012 – November 10, 2012 – p.2
Optical Network Design
Optical networks: the foundation of the global network infrastucture
Network design and planning crucial to operation of the Internet:
QoS, support of critical applications
survivability to failures
economics
· · ·
ACP 2012 – November 10, 2012 – p.3
Challenges
Network design problems are hard
Optimal solutions do not scale with
network size
number of wavelengths (≈ 100/fiber currently)
ACP 2012 – November 10, 2012 – p.4
Challenges
Network design problems are hard
Optimal solutions do not scale with
network size
number of wavelengths (≈ 100/fiber currently)
“What-if” analysis: substantial investments to explore sensitivity to:
forecast traffic demands
capital/operational cost assumptions
service price structures
· · ·
ACP 2012 – November 10, 2012 – p.4
Traffic Grooming as Optimization Problem
Inputs to the problem:
physical network topology (fiber layout)
number of wavelengths W and their capacity C
traffic matrix T = [tsd] → int multiples of unit rate (e.g., OC-3)
Output:
logical topology
traffic grooming on lightpaths
lightpath routing and wavelength assignment (RWA)
Objective:
minimize the number of lightpaths so as to carry the traffic
ACP 2012 – November 10, 2012 – p.6
Traffic Grooming Subproblems
1
2
3
4
5
6
Logical topology design → determine the lightpaths to be established
ACP 2012 – November 10, 2012 – p.7
Traffic Grooming Subproblems
1
2
3
4
5
6
Logical topology design → determine the lightpaths to be established
Traffic routing → route traffic on virtual topology
ACP 2012 – November 10, 2012 – p.7
Traffic Grooming Subproblems
1
2
3
4
5
6
Logical topology design → determine the lightpaths to be established
Traffic routing → route traffic on virtual topology
Lightpath routing → route the lightpaths over the physical topology
ACP 2012 – November 10, 2012 – p.7
Traffic Grooming Subproblems
1
2
3
4
5
6
Logical topology design → determine the lightpaths to be established
Traffic routing → route traffic on virtual topology
Lightpath routing → route the lightpaths over the physical topology
Wavelength assignment → assign wavelengths to lightpaths w/o clash
ACP 2012 – November 10, 2012 – p.7
Airline Analogy (2)
Lightpaths, logical topology ↔ Flights, flight routes
Traffic routing ↔ Travel itinerary
Electronic ports ↔ Gates
Grooming switch ↔ Hub airport
Wavelengths ↔ Gate timeslots
ACP 2012 – November 10, 2012 – p.8
Traffic Grooming Complexity
. . . . . .S 1 2 3 n n+1 n+2 n+3 n+4 2n+1 D
Problem instance:
unidirectional linear (path) network
logical topology and RWA is given
traffic either bifurcated or not bifurcated
Objective: find a routing of traffic onto the lightpaths
Result: problem is NP-complete → reduction from Subset Sums
ACP 2012 – November 10, 2012 – p.9
Challenge: Running Time
4 6 8 10 120.01
0.1
1
10
100
1000
10000
100000
360000
tLim
N
Run
ning
tim
e (s
ec)
Original
ACP 2012 – November 10, 2012 – p.10
Challenge: Wavelength Fragmentation
1 2 3 4 5
2
4
6
8
10
12
14
16
18
20
22
Traffic Instance
Num
ber
of w
avel
engt
hs in
sol
utio
n
DecompositionOriginal, W
a=5
Original, Wa=10
Original, Wa=30
ACP 2012 – November 10, 2012 – p.11
Routing and Wavelength Assignment (RWA)
Fundamental control problem in optical networks
Objective: for each connection request determine a lightpath, i.e.,
a path through the network, and
a wavelength
Two variants:
1. online: lightpath requests arrive/depart dynamically
2. offline: set of lightpaths to be established simultaneously
ACP 2012 – November 10, 2012 – p.12
Offline RWA
Input:
network topology graph G = (V,E)
traffic demand matrix T = [tsd]
Objective:
establish all lightpaths with the minimum # of λs
maximize established lightpaths for a given # of λs
Constraints:
each lightpath uses the same λ along path
lightpaths on same link assigned distinct λs
NP-hard problem (both objectives)
ACP 2012 – November 10, 2012 – p.13
Ring RWA: Running Time
6 8 10 12 14 16 18 20 22 24<0.001
0.01
0.1
1
10
100
1000
7200tLim
Mem
N
sol t
ime
(s)
linkpathMIS
ACP 2012 – November 10, 2012 – p.16
Ring RWA: Running Time
6 8 10 12 14 16 18 20 22 24<0.001
0.01
0.1
1
10
100
1000
7200tLim
Mem
N
sol t
ime
(s)
linkpathMISMISD−2
ACP 2012 – November 10, 2012 – p.16
Ring RWA: Running Time
6 8 10 12 14 16 18 20 22 24<0.001
0.01
0.1
1
10
100
1000
7200tLim
Mem
N
sol t
ime
(s)
linkpathMISMISD−2MISD−4
ACP 2012 – November 10, 2012 – p.16
Ring RWA: Running Time
6 8 10 12 14 16 18 20 22 24<0.001
0.01
0.1
1
10
100
1000
7200tLim
Mem
N
sol t
ime
(s)
linkpathMISMISD−2MISD−4MISD−8
ACP 2012 – November 10, 2012 – p.16
Mesh RWA
Path formulation:
compact formulation for optimal symmetric solutions
fast, close to overall optimal
ACP 2012 – November 10, 2012 – p.17
Symmetric Solution: Running Time
NSF (14) German (17) EON (20) USA (32)1
10
100
1000
10000
72000
tLim
Network Topologies (Number of Nodes)
Sol
utio
n T
ime
(S
ec)
Original ILP, Asymmetric TrafficOriginal ILP, Symmetric TrafficNew ILP, Asymmetric TrafficNew ILP, Symmetric Traffic
ACP 2012 – November 10, 2012 – p.18
Symmetric Solution: Quality
2 4 6 8
10
20
30
40
50
60
70
Tmax
Sol
utio
n V
alue
, W
LFAP Heuristic, Asymmetric TrafficLFAP Heuristic, Symmetric TrafficOriginal ILP, Asymmetric TrafficNew ILP, Asymmetric TrafficOriginal ILP, Symmetric TrafficNew ILP, Symmetric Traffic
ACP 2012 – November 10, 2012 – p.19
Traffic Grooming: Integrate MISD
4 6 8 10 120.01
0.1
1
10
100
1000
10000
100000
360000
tLim
N
Run
ning
tim
e (s
ec)
Originalw/ MISD
ACP 2012 – November 10, 2012 – p.20
Traffic Grooming Decomposition
Decompose and solve the two problems sequentially:
1. Logical topology and traffic routing
2. Routing and wavelength assignment
ACP 2012 – November 10, 2012 – p.21
Traffic Grooming Decomposition
Decompose and solve the two problems sequentially:
1. Logical topology and traffic routing→ determine lightpaths→≤ objective value of integrated problem
2. Routing and wavelength assignment
ACP 2012 – November 10, 2012 – p.21
Traffic Grooming Decomposition
Decompose and solve the two problems sequentially:
1. Logical topology and traffic routing→ determine lightpaths→≤ objective value of integrated problem
2. Routing and wavelength assignmentroute and color lightpaths from Step 1fast (for rings and medium mesh networks)
ACP 2012 – November 10, 2012 – p.21
Traffic Grooming Decomposition
Decompose and solve the two problems sequentially:
1. Logical topology and traffic routing→ determine lightpaths→≤ objective value of integrated problem
2. Routing and wavelength assignmentroute and color lightpaths from Step 1fast (for rings and medium mesh networks)
Optimal for instances thatare not λ-limited
1 2 3 4 5
2
4
6
8
10
12
14
16
18
20
22
Traffic Instance
Num
ber
of w
avel
engt
hs in
sol
utio
n
DecompositionOriginal, W
a=5
Original, Wa=10
Original, Wa=30
ACP 2012 – November 10, 2012 – p.21
Logical Topology and Traffic Routing Problem
i jb integer
ij
Independent of physical topology
Integer variables are not binary→ LP relaxation possible
ACP 2012 – November 10, 2012 – p.22
Iterative Algorithm
1. thresh ← 0
2. Relax integrality constraints on lightpath variables s.t.:bij − ⌊bij⌋ > thresh
3. Solve relaxed problem
4. If all variables integer, stop
5. If thresh > .8, stop
6. thresh + = 1/C
7. Repeat from Step 2
ACP 2012 – November 10, 2012 – p.23
Iterative Algorithm: Quality
100
110
120
130
140
0 0.2 0.4 0.6 0.8 1
% fr
om o
ptim
al
thresh
N = 8N = 16N = 24N = 32
ACP 2012 – November 10, 2012 – p.24
Iterative Algorithm: Running Time
0.01
0.1
1
10
100
1000
10000
50000
8 16 24 32
Run
ning
tim
e (s
ec)
N
thresh = 0.8
ACP 2012 – November 10, 2012 – p.25
Conclusion & Ongoing Research
First steps towards efficient network design
scalable techniques on commodity hardware
lower the barrier to entry
focus on exploring design options, not ILP details
ACP 2012 – November 10, 2012 – p.26
Conclusion & Ongoing Research
First steps towards efficient network design
scalable techniques on commodity hardware
lower the barrier to entry
focus on exploring design options, not ILP details
Many open problems:
impairment-aware RWA
shared protection, survivable grooming
routing and spectrum allocation in elastic optical networks
ACP 2012 – November 10, 2012 – p.26