Transcript Power-Aware Network Design

POWER-AWARE
NETWORK DESIGN
«Power Awareness in Network Design and Routing»
J. Chabarek et al.
«Energy-Minimized Design for IP Over WDM Networks»
G. Shen, R. S. Tucker
Introduction
• The Internet is expanding tremendously
• Growth in the number of end users and connection speeds ->
exponential increase in bandwidth demand
• Increase in energy consumption
• Cost of transmission and switching one of the major
barriers
• Energy consumption may become a barrier
• 1% - 2% of total electricity consumption in US
• A cut of 1% in the Internet energy consumption means about US$5 billion per year
• Increase in power density
• Thermal issues -> limitations of air cooling
• Increase in operational costs
• Increase greenhouse footprint
• Save the Earth!!!
Power Aware Design Areas (I)
• Three main areas for power aware design
• System Design
• Development in CMOS technology -> improvements are slowing down
• Multi-Chassis Systems: separate physical components clustered
forming a single logical router
• Aggregate power consumption increases -> heat spread over a large physical area ->
existing cooling techniques used
• Alternative Systems: optical switches
• Terabits of bandwidth at much lower power dissipation
• Protocols
• Investigated in wireless networks -> Opportunities in wire-line networks
• Basic notion: put components to sleep if low traffic load
• Routing protocols: routes calculated with power consumption
constraints
Power Aware Design Areas (II)
• Network Design
• Deploy routers such that the aggregate power demand is minimized
• Satisfying robustness and performance
• Two approaches
• Multiple router-level topologies satisfying capacity, robustness and power
consumption
• Limit power-hungry systems to a subset of routers
• Selection of chassis and line cards in routers is a main issue to reduce
power consumption
• In IP over WDM networks
• IP routers use more than 90% of total power
• Lightpath bypass is used to reduce the number of IP router ports -> IP ports
consume major energy in IP routers
Router Power Consumption
• Router power consumption depends on
• Type of router chassis
• Type and number of line cards deployed in the chassis
• Configuration and operating conditions
• Size of packets
• 100 bytes / 576 bytes / 1500 bytes
• Size of forwarding table
• 1000 entries / 32000 entries
• Type of traffic
• UDP
• TCP
• Employed protocols and techniques
•
•
•
•
•
OSPF
Netflow
Unicast Reverse Path Forwarding (uRPF)
Access Control List (ACL)
Active Queue Management - Random Early Detection (AQM – RED)
Router Power Consumption
• Chassis and line card combinations
• Chassis: Cisco GSR 12008 / Cisco 7507
Router Power Consumption
• Chassis and line card combinations (cont.)
• Base system is the most consuming
• 7507 chassis + router processor -> 210 Watts
• GSR chassis + router processor + switching fabric -> 430 Watts
 It is best to minimize the number of chassis and maximize the
number of line cards per chassis
• Calculated power consumption of different cards
Router Power Consumption
• Configuration and operating conditions
• A 4-port Gigabit Ethernet line card and a OC-48 card in a GSR chassis
is used
• Deployed testbed:
Router Power Consumption
• Configuration and operating conditions (cont.)
• Constant bit rate UDP traffic and different packet sizes
• 1500 bytes / 576 bytes / 100 bytes
 Power consumption increases as packets get smaller!!!
Router Power Consumption
• Configuration and operating conditions (cont.)
• Constant bit rate UDP traffic, medium packets and different features
• Large forwarding table / ACL / uRPF / OSPF
 uRPF is the most consuming
 Large forwarding table is less consuming!!
Router Power Consumption
• Configuration and operating conditions (cont.)
• Self-similar TCP traffic, 75% offered load and different features
• Netflow / AQM - RED
 Power consumption similar to UDP with large-sized packets
Router Power Consumption
• Configuration and operating conditions (cont.)
• Maximum variation in previous slides -> 20 Watts
• Extrapolating a fully loaded chassis -> 150 - 200 Watts
• Less significant than chassis/line card configuration
• General Model:
• PC -> power consumption of router
• X is a vector defining chassis type, line cards, configuration and traffic profile
•
•
•
•
CC -> power consumption of a chassis type
N -> number of line cards
TP -> scaling factor (traffic utilization)
LCC -> cost of line card
Power Consumption Optimization
• Main focus: allocation of line cards and chassis in nodes
to minimize power consumption
• Mixed-Integer resource allocation problem with
multicommodity flow constraints
• Inputs
• Network with OSPF link weights
• Traffic matrix
• Line card and chassis options
• Outputs
• How each node should be provisioned
• Multipath routing
• Implemented with General Algebraic Modeling System
(GAMS)
Power Consumption Optimization
• Networks are taken from the Rocketfuel project
• Inferred weights and link latencies
• Link weights -> calculate approximate bandwidths of each link
• Traffic matrixes generated with a gravity model
• Three additional random graphs with 12 nodes and varying number of
directed edges (Waxman method)
Power Consumption Optimization
• Network design problem: deploy different chassis/line
card configurations such that provisioning requirements
are satisfied and power consumption in minimized
• Traffic is scaled for each origin-destination pair -> linear scaling factor
• Varies provisioning requirements
• Traffic flows might be altered to put cards/chassis to sleep in low utilization
• First scenario includes only GSR chassis and OC-48 line card
• Only 10 line cards allowed per chassis
• Scaling factor varies from 0.1 to 40
Power Consumption Optimization
• Other experiments relaxing line cards per chassis, chassis type and
card types (not in the paper)
 Minimum power consumption -> chassis accommodating large numbers
of line cards and line cards capacities that closely match demand
Power Consumption Optimization
• Power savings
• Compared to a non-power-aware network design (shortest path)
• Using a specific chassis (GSR) and line cards (OC-48 or 0C-12)
• OC-12 line cards achieve smaller savings -> more ingress/egress node
ports
• Cost of additional connectivity is zero as long as the number of ports does not require
additional line cards
IP Over WDM Network
• IP layer:
• Core IP router aggregates data traffic from low-end access routers
• IP router ports consume major energy (forwarding process) -> number of IP ports as
measure of total power consumption
• Optical layer
• Optical switches interconnected with physical fiber links
• May contain multiple fibers
• Each fiber needs a pair of multiplexer/demultiplexer
• Each wavelength require a pair of transponders -> full wavelength conversion is
assumed
• EDFA amplifiers are deployed on fiber links
IP Over WDM Network
• Two implementation approaches
• Lightpath non-bypass
• All data carried by lightpaths is processed and forwarded by IP routers
• All lightpaths incident to a node must be terminated
• Lightpath bypass
• IP traffic whose destination is not the intermediate node -> directly
bypasses the intermediate router
• Saves IP router ports
Energy Consumption Optimization for IP
over WDM
• Network design problem: design an energy-minimized IP
over WDM network
• Serving all the traffic demands
• With a limited maximal number of wavelengths in each fiber
• With a limited maximal number of IP router ports at each node
• Inputs
• Physical topology -> N nodes and E links
• Traffic demand matrix
• Number of wavelength channels per fiber and capacity of each wavelength
• Maximal number of IP router ports at each node
• Energy consumption of router ports, transponders and EDFAs
Energy Consumption Optimization for IP
over WDM
• The optimization problem is solved using a Mixed-Integer
Linear Programming (MILP) model including
• Energy consumption of IP routers, EDFAs and transponders
• Layout of EDFAs
• Ports for aggregating data from low-end routers
• MILP model minimizes also the number of network
components -> could be used for cost-minimized IP over
WDM network
• The computational complexity is high
• O(N4) variables and O(N3) constraints
• Heuristics are needed for fast solution
Energy Consumption Optimization for IP
over WDM - Heuristics
• Heuristics
• Direct Bypass: directly establish virtual links (lightpaths) whose capacity
is sufficient to accommodate all the traffic demands between each node
pair
• Routing of lightpaths -> shortest path routing
• Simple
• Could lead to low capacity utilization
• Multi-hop bypass: traffic demands between different node pairs could
share capacity on common lightpaths
• Elongate lengths of some IP traffic flows
• Fewer lightpaths -> fewer IP router ports
Energy Consumption Optimization for IP
over WDM - Heuristics
• Multi-hop bypass heuristic:
Energy Consumption Optimization for IP
over WDM - Setup
• Five study cases
• Linear relaxation of the MILP model -> lower bound
• MILP optimal design
• Non-bypass -> upper bound
• Direct bypass
• Multi-hop bypass
• Inputs
• Traffic demand between
each pair node:
• Uniform distribution
within a certain range
centered at an identical
average
Energy Consumption Optimization for IP
over WDM – Test Networks
• Test networks
n6s8
NSFNET
USNET
Energy Consumption Optimization for IP
over WDM – Total Power Consumption
n6s8
NSFNET
Larger topology ->
higher power consumption,
heuristics closer to lower bound
Non bypass -> upper bound
LP relax. -> lower bound
USNET
Linear relationship between total
power consumption and total traffic
demand intensity
Energy Consumption Optimization for IP
over WDM – Power Consumption Saving
n6s8
NSFNET
Larger topology -> higher savings,
longer lightpaths bypassing more
nodes -> fewer IP ports
Multi-hop bypass heuristic performs
better than direct bypass ->
Small traffic flows are aggregated
USNET
Energy Consumption Optimization for IP
over WDM – Component Consumption
n6s8
NSFNET
Energy Consumption Optimization for IP
over WDM – Geographical Distribution
n6s8
NSFNET
All bypass design have a more uniform power distribution
Solve problems associated with:
Supplying large amounts of energy
Removing associated heat
Energy Consumption Optimization for IP
over WDM – Cost Analysis
• The model could be used for minimizing cost
• Changing the optimization weights from energy to cost
• May NOT be valid if components with low energy consumption are the
most expensive ones
N6s8 network based on the
MILP optimization model
Conclusions
• Energy consumption may become a barrier for the
Internet
•
•
•
•
Operational costs
Greenhouse footprint
Cooling issues
Supplying large amounts of energy
• Power aware design could solve it
• Power aware system design
• Power aware protocols
• Power aware network design
• Power aware network design could achieve important
savings
• In IP over WDM networks, lightpath bypass could save power
consumption
References
• [CHA08] J. Chabarek et al., «Power Awareness in
Network Design and Routing», Proc. Of IEEE INFOCOM,
2008
• [SHE09] G. Shen, R. S. Tucker, «Energy-Minimized
Design for IP Over WDM Networks», Journal of Optical
Communication Networks, June 2009.