Power Aware and Computationally EfficientOptical Network Design
George N. Rouskas
Department of Computer Science
North Carolina State University
Joint work with: Dr. Emre Yetginer (Tubitak, Turkey), Zeyu Liu (NCSU)
IEEE DLT, May 2011 – p.1
Outline
Power-Aware Traffic Grooming
Power Consumption in Networks: Trends and Challenges
Optical Networks to the Rescue: Power-Aware Traffic Grooming
Results and Discussion
Computationally Scalable Optical Network Design
Routing and Wavelength Assignment (RWA)
New Computationally Efficient ILP Formulations for Ring and Mesh
Numerical Results
Conclusions and Future Research Directions
IEEE DLT, May 2011 – p.2
The Challenge of Power Consumption
Power consumption a growing challenge for ICT industry:
high operating costs
high capital costs→ cooling equipment
Significant environmental impact
industry responsible for≈2-3% of man-made CO2
growing at double-digit rates
IEEE DLT, May 2011 – p.3
Why Energy Efficiency For Networks
So far, energy efficiency focus has been on servers and cooling
Networks are shared resources→ always on
In the US: 6 TWatts of power on networks
IEEE DLT, May 2011 – p.4
Addressing the Challenge
Energy-efficient designs:
1. low-power techniques in design of componentssupport low-power states in processors, memory, disksdisable clock signal to unused parts of processorreplace complex uniprocessors with multiple simple cores
2. power management techniques across systemsintelligent policies to exploit low-power statesworkload management
IEEE DLT, May 2011 – p.5
Addressing the Challenge
Energy-efficient designs:
1. low-power techniques in design of componentssupport low-power states in processors, memory, disksdisable clock signal to unused parts of processorreplace complex uniprocessors with multiple simple cores
2. power management techniques across systemsintelligent policies to exploit low-power statesworkload management
Seek inexpensive energy sources
→ build data/compute centers wherever energy is cheap
IEEE DLT, May 2011 – p.5
The Networking Infrastructure
Forwarding table lookup→ routers operate at very high speeds
high energy consumption
low-power operation not feasible
IEEE DLT, May 2011 – p.6
The Networking Infrastructure
Forwarding table lookup→ routers operate at very high speeds
high energy consumption
low-power operation not feasible
New routing architecture?
partition Internet address space
multiple parallel networks of “virtual” routers
each network handles small address space→ energy-efficientrouters
IEEE DLT, May 2011 – p.6
Trends: Traffic vs. Router Capacity Growth
Internet traffic is already in the Exabytes (10 bytes) per-year, and is estimated to reach Z
with
has been do
past 20
has been doubling
months
forward
lag
speeds
unless som
operators wil
routers as today
times the space, 256 times the power,
and cost 100 times as much
IEEE DLT, May 2011 – p.7
Trends: Energy Demand Will Exceed Supply
Average access rate [Mbit/s]
Pow
er
[MW
/10
6users
]
Metro +Access
Core
WDM links
Total
0
100
200
300
0 200 400 600 800 10000
2
4
6
10
12
% o
f ele
ctr
icity s
upply
8
14Dominated bycore routers
of the world’s population were to obtain broadband access:
IEEE DLT, May 2011 – p.9
Trends: Energy Demand Will Exceed Supply
Average access rate [Mbit/s]
Pow
er
[MW
/10
6users
]
Metro +Access
Core
WDM links
Total
0
100
200
300
0 200 400 600 800 10000
2
4
6
10
12
% o
f ele
ctr
icity s
upply
8
14Dominated bycore routers
of the world’s population were to obtain broadband access:If 33% of the world’s population were to obtain broadband access:
Access rate 1 Mbps 10 Mbps
Power consumption 100 GW 1 Tw
electricity supply 5% 50%
IEEE DLT, May 2011 – p.9
Optical Networks to the Rescue
Optical networks:
energy efficientmany passive componentsactive components (e.g., repeaters) can be solar/wind-powered
low carbon footprint
IEEE DLT, May 2011 – p.10
Motivation: Router Power Consumption
Juniper Core Router T640
8 ports at 40 Gbps each
Power consumption:
4500 W overall, 550 W/port
Cost (10c/kWh):
$4000/year, $500/port/year
Add AC+UPS:
≈ double power consumption→$1000/port/year
Power consumption increases with line rate
IEEE DLT, May 2011 – p.11
Motivation: Optical Switch Power Consumption
Calient DiamondWave PXC 128
128×128 switch
Power Consumption:
< 750 W overall
< 6 W/port
independent of line rate
PXC consumes ≈ 1% of powerper port consumed by the Juniperrouter
IEEE DLT, May 2011 – p.12
The Case for Optical Bypass
26
2
3
45
6
7
8
9
10
11
12
1314
15
16
17
18
19
20
21
22
23
24
31
32
30
28 29
25
27
1
Most (≈ 80%) network links: < 200 miles in length
Most traffic demands (≈ 80%): travel > 200 miles
IEEE DLT, May 2011 – p.13
The Case for Optical Bypass
26
2
3
45
6
7
8
9
10
11
12
1314
15
16
17
18
19
20
21
22
23
24
31
32
30
28 29
25
27
1
Most (≈ 80%) network links: < 200 miles in length
Most traffic demands (≈ 80%): travel > 200 miles
IEEE DLT, May 2011 – p.13
Grooming Networks
DXC/Router
OXC
Node 1
Node 2
Node 3
Node 4
Node 5
Node 6
Electronic LayerVirtual TopologyCapacity Constraint
Optical LayerPhysical TopologyWavelength Constraint
Data Flow From Node1 to Node 6
Data Flow From Node3 to Node 6
Optical Fiber
What is traffic grooming?
Efficiently set up lightpaths and groom (i.e., pack/unpack,switch, route, etc.) low-speed traffic onto high capacitywavelengths so as to minimize network resources
IEEE DLT, May 2011 – p.14
Traffic Grooming as Optimization Problem
Inputs to the problem:
physical network topology (fiber layout)
traffic matrix T = [tsd]→ int multiples of unit rate (e.g., OC-3)
Output:
logical topology
lightpath routing and wavelength assignment (RWA)
traffic grooming on lightpaths
IEEE DLT, May 2011 – p.15
Traffic Grooming Subproblems
1
2
3
4
5
6
Logical topology design→ determine the lightpaths to be established
IEEE DLT, May 2011 – p.16
Traffic Grooming Subproblems
1
2
3
4
5
6
Logical topology design→ determine the lightpaths to be established
Lightpath routing→ route the lightpaths over the physical topology
IEEE DLT, May 2011 – p.16
Traffic Grooming Subproblems
1
2
3
4
5
6
Logical topology design→ determine the lightpaths to be established
Lightpath routing→ route the lightpaths over the physical topology
Wavelength assignment→ assign wavelengths to lightpaths w/o clash
IEEE DLT, May 2011 – p.16
Traffic Grooming Subproblems
1
2
3
4
5
6
Logical topology design→ determine the lightpaths to be established
Lightpath routing→ route the lightpaths over the physical topology
Wavelength assignment→ assign wavelengths to lightpaths w/o clash
Traffic grooming→ route traffic on virtual topology
IEEE DLT, May 2011 – p.16
Grooming Objectives
Minimize the number of lightpaths→ minL
equivalent to minimizing the number of electronic ports
minimizes initial deployment cost
IEEE DLT, May 2011 – p.17
Grooming Objectives
Minimize the number of lightpaths→ minL
equivalent to minimizing the number of electronic ports
minimizes initial deployment cost
Minimize the amount of electronically switched traffic→ minT
minimizes average processing delay
minimizes electronic switching capacity
IEEE DLT, May 2011 – p.17
Grooming Objectives
Minimize the number of lightpaths→ minL
equivalent to minimizing the number of electronic ports
minimizes initial deployment cost
Minimize the amount of electronically switched traffic→ minT
minimizes average processing delay
minimizes electronic switching capacity
Minimize the amount of power consumption→ minP
maximizes power efficiency (in Watts/bit)
minimizes operational costs
most general objective
IEEE DLT, May 2011 – p.17
Power Consumption: Assumptions
Optical power≪ Electronic power→ energy consumed by optical ports is negligible
Inactive ports and transceivers may be shut down
Power consumption of each component (electronic input/output port,O/E and E/O converters) increases linearly with amount of traffichandled
IEEE DLT, May 2011 – p.18
Power Consumption: Router Port Model
pow
er, P
P0
maxP
C
p=0
traffic rate, t
Equivalent to minimizing the number of lightpaths→ minL
IEEE DLT, May 2011 – p.19
Power Consumption: Router Port Model
pow
er, P
P0
maxP
C
P =00
traffic rate, t
Equivalent to minimizing amount of electronically switched traffic→ minT
IEEE DLT, May 2011 – p.19
Power Consumption: Router Port Model
P=P +pt
P0
maxP
C
p=0
P =00
traffic rate, t
pow
er, P 0
Most general model: minimize power consumption→ minP
IEEE DLT, May 2011 – p.19
ILP Formulation
Objective:
1. minL: min # of lightpaths
2. minT: min amount of electronically switched traffic
3. minP: min power consumption→ most general model
subject to:
lightpath routing constraints
wavelength assignment constraints
traffic routing constraints
IEEE DLT, May 2011 – p.20
Performance Evaluation
6
1
2 3
4
5
W = 3 wavelengths
C = 48 wavelength capacity
Source-destination traffic tsd ← uniform[0, tmax]
P0 = 0.25
Pmax = 1
Each data point: average of 40 problem instances
IEEE DLT, May 2011 – p.21
Results: Number of Lightpaths
0 10 20 30 40 50 605
10
15
20
25
30
# lig
htpa
ths
tmax
minTminLminP
IEEE DLT, May 2011 – p.22
Results: Amount of Electronically Switched Traffic
0 10 20 30 40 50 600
50
100
150
200
amou
nt o
f tra
ffic
switc
hed
tmax
minTminLminP
IEEE DLT, May 2011 – p.23
Results: Relative Power Consumption
0 10 20 30 40 50 600
10
20
30
40
50
60
70
80
exce
ss p
ower
con
sum
ptio
n w
.r.t.
min
P (
%)
tmax
minTminL
IEEE DLT, May 2011 – p.24
Summary and Discussion
Power-aware design may lead to significant energy savings even forsmall networks
The benefits are expected to increase with the network size
Challenges:
existing ILPs do not scale to realistic networks
performance of heuristics difficult to characterize
IEEE DLT, May 2011 – p.25
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 RWA: connection requests arrive/depart dynamically
2. static RWA: a set of traffic demands to be establishedsimultaneously
IEEE DLT, May 2011 – p.26
Static RWA
Input:
network topology graph G = (V,E)
traffic demand matrix T = [tsd]
Objective:
minRWA: establish all demands with the minimum # of λs
maxRWA: maximize established demands for a given # of λs
Constraints:
wavelength continuity: each lightpath uses the same λ along path
distinct wavelength: lightpaths using the same link assigneddistinct λs
NP-hard problem (both variants)
IEEE DLT, May 2011 – p.27
Link ILP Formulation
Nodes/links are entities of interest
Focus on traffic demand to and from nodes, on links
lw
ij
link l
c =0,1ij
Bridging variable: demand between nodes on links
IEEE DLT, May 2011 – p.30
Path ILP Formulation
Nodes/paths are entities of interest
Demand is still between nodes
For each given demand node pair, list all paths→ typically, a subset of all paths
...
i j
k
1
2
assign variable to path traffic flow→ implicitly identifies demand
for each link, sum up path flow variables→ constrain with capacities
IEEE DLT, May 2011 – p.31
Maximal Independent Sets
Independent set: a set of vertices in a graph no two of which areadjacent
Maximal independent set: not a subset of any other independent set
6
1 2
3 4
5 6
1 2
3 4
5 6
1 2
3 4
5 6
1 2
3 4
5 6
1 2
3 4
5
IEEE DLT, May 2011 – p.33
MIS ILP Formulation
Precompute k paths for each source-destination pair
Create the path graph Gp:
each node in Gp corresponds to a path in the original network
two nodes connected in Gp if corresponding paths share a link
Enumerate the MISs of Gp
Set up ILP to assign wavelengths to each MIS
IEEE DLT, May 2011 – p.34
Comparison
Link ILP formulation
O(N4W ) variables
O(N3W ) constraints
symmetry with respect to λ permutations
Path ILP formulation
O(N2W ) variables
O(N2W ) constraints
symmetry with respect to λ permutations
MIS ILP formulation
O(3N2/3) variables
O(N2) constraints
no symmetry
size independent of W → future-proof
IEEE DLT, May 2011 – p.35
RWA in Ring Networks
Vast parts of network infrastructure based on SONET/SDH rings
AT&T operates≈ 6,700 rings in North America
→ optimal solutions for rings important for foreseeable future
Max size of SONET ring: 16 nodes
Operators have started transition to mesh networks→ next ...
IEEE DLT, May 2011 – p.36
Running Time Results, W = 120
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)
link
IEEE DLT, May 2011 – p.37
Running Time Results, W = 120
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)
linkpath
IEEE DLT, May 2011 – p.37
Running Time Results, W = 120
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
IEEE DLT, May 2011 – p.37
MIS Decomposition for Rings: MISD-2
Clockwise paths do not intersect with counter-clockwise paths:
Gp = Gcwp ∪Gccw
p
M,M cw,M ccw : # of MISs of Gp, Gcwp , Gccw
p :
M cw = M ccw =√
M
→ orders of magnitude decrease in # of variables/size of formulation
Slight modifications to formulation
IEEE DLT, May 2011 – p.38
Further Decomposition: MISD-4
Consider clockwise direction only→ similar steps for counter-clockwise
Partition ring in two parts such that:
Gcwp = Gcw,0
p ∪Gcw,1p ∪Gcw,core
p
cw,0
1
2
3
4
1
2
3
4
Gp
Gpcw,core
cw,1
Gp
IEEE DLT, May 2011 – p.39
MISD-4 (cont’d)
Express each MIS m of Gcwp as:
m = m0 ∪m1 ∪ q
Modify the formulation appropriately
# MIS variables ↓# constraints ↑
Recursively partition the two ring parts to effect higher-orderdecompositions (MISD-8, MISD-16, . . .)
IEEE DLT, May 2011 – p.40
Results: # of MIS Variables
6 8 10 12 14 16 18 20 22 24 26 28 30 3210
1
102
103
104
105
106
107
108
109
1010
N
# of
MIS
s
MISMISD−2MISD−4MISD−8
IEEE DLT, May 2011 – p.41
Running Time Results, W = 120
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
IEEE DLT, May 2011 – p.42
Running Time Results, W = 120
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
IEEE DLT, May 2011 – p.42
Running Time Results, W = 120
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
IEEE DLT, May 2011 – p.42
Running Time Results, W = 120
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
IEEE DLT, May 2011 – p.42
RWA in Mesh Networks
MIS decomposition does not work
Devised new exact decompositions for path formulation
May solve efficiently 40-node networks
IEEE DLT, May 2011 – p.43
Running Time Results: Torus
9 16 25 360.01
0.1
1
10
100
1000
10000
72000
tLim
Num of Nodes
Sol
utio
n T
ime
(s)
Original Path
IEEE DLT, May 2011 – p.44
Running Time Results: Torus
9 16 25 360.01
0.1
1
10
100
1000
10000
72000
tLim
Num of Nodes
Sol
utio
n T
ime
(s)
Original PathImproved Path
IEEE DLT, May 2011 – p.44
Running Time Results: Asymmetric Topologies
NSF(14) Havana(17) EON(20) US 32−node1
10
100
1000
10000
72000
tLim
Networks(Number of Nodes)
Sol
utio
n T
ime
(s)
Original Path
IEEE DLT, May 2011 – p.45
Running Time Results: Asymmetric Topologies
NSF(14) Havana(17) EON(20) US 32−node1
10
100
1000
10000
72000
tLim
Networks(Number of Nodes)
Sol
utio
n T
ime
(s)
Original PathImproved Path
IEEE DLT, May 2011 – p.45
Conclusion & Ongoing Research
Traffic grooming is ideal candidate for encompassing energy concerns
Power-aware network design may lead to significant energy savings
RWA subproblem can be solved efficiently
Current research focuses on:
more accurate power consumption models for traffic grooming
computationally efficient formulations for optical network designproblems
traffic groomingimpairment-aware RWAmulticast RWA and grooming
IEEE DLT, May 2011 – p.46
Top Related