Green Domino Incentives:
Impact of Energy-aware
Adaptive Link Rate
Policies in Routers
Cyriac James
University of Calgary
Canada
Niklas Carlsson
Linköping University
Sweden
Presented by Martin Arlitt, HP Labs
2
Motivation
—  Energy savings in Internet routers
—  Over-provisioned to meet peak traffic
—  Hence, often under utilized
—  Effect on downstream routers
—  Positive or negative
—  Energy and Delay
3
Contribution
—  Evaluation Framework
—
—
—
—
Router Model
Policy Model
Energy Model
Traffic Model
—  Trace based simulation
—  Capture real traffic characteristics
—  Analysis on immediate downstream router
—  Delay
—  Improvement in energy savings
4
Adaptive Link Rate (ALR)
—  Energy saving techniques
—  Rate scaling
—  Active/idle toggling
—  IEEE 802.3az
—  Commercial
—  Cisco Catalyst 4500E Switch
48-port Line Card (Photo Courtesy: Cisco)
Symbolic representation of port operation
5
Policy Parameters & Delay
2
70
Per Router Packet Delay (ms)
Per Router Packet Delay (ms)
10
1
10
0
10
−1
10
−2
10
0
100
200
300
400
500
600
Active Link Rate (Mbps)
Rate Scaling
700
800
900
1000
60
50
40
30
20
10
0
2
3
10
10
4
10
5
10
Threshold Value (bytes)
Active/Idle Toggling
•  Rate scaling
•  Service rate or port speed
•  Reduction in speed
Energy Savings
•  Active/Idle Toggling
•  Queue threshold
•  Amount of idle time
Energy Savings
6
Evaluation Framework
Router
Model
Policy
Model
Design
Issues
Energy
Model
Traffic
Model
7
Policy Model
—  Tail delay (99th percentile)
—  Between .01ms and 100ms
—  Vary policy parameters
—  Port rate
—  Queue threshold
—  Hybrid
—  Port rate
—  Queue threshold < Smallest packet
8
Router Model
—  Delay
—  Switch Fabric
—  Queue
—  Transmit
—  Model by Hohn et al. 2009
—  Switch fabric delay: 10 – 50 microseconds
—  Delay constraints in milliseconds
—  Delay = Queue delay + Transmit delay
—  Infinite queue
—  Tail delay
Router
9
Energy Model
—  Proportional Model
—  Interested in
Relative energy consumption
—  NOT absolute
—  Relative increase/decrease in energy savings
—  At R2, R3 and R4
—  R1 runs green techniques
—  R1 does not
10
Traffic Model
—  Traffic scenarios
—  Dispersion: 1*2
—  Aggregation: 2*1
—  Multiplexing: 1*1, 2*2 (shown), 3*3
—  Packet traces (public)
—  Waikato trace (edge)
—  MAWI (core)
11
Simple Back-to-Back Case
—  Past studies on tandem queues
—  Increased delay at R2 for (utilization < 60%)
—  Continuous and independent service time
—  Our results:
1
0.9
0.8
Empirical CDF
0.7
R1 (1.2%)
0.6
R2 (1.2%)
R1 (15%)
0.5
R2 (15%)
0.4
0.3
0.2
0.1
0 −4
10
−3
10
−2
10
−1
10
0
10
Per Router Packet Delay (ms)
1
10
2
10
12
Bimodal Distribution
1
Edge: Outgoing
Core: Direction−B
Core: Direction−A
Edge: Incoming
0.9
0.8
Empirical CDF
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
200
400
600
800
1000
1200
1400
1600
Packet Size (bytes)
•  Most packet sizes are either small (<100 bytes) or large (>1400
bytes)
•  Incoming edge traffic has more large packets
13
Back-to-Back Probability
Small:<= 100 byes
Large: >=1400 bytes
Medium: > 100 and < 1400
Small Medium Large
Small
Small Medium Large
0.39
0.11
0.04
Small
0.23
0.05
0.07
Medium 0.10
0.06
0.03
Medium 0.04
0.02
0.04
Large
0.02
0.20
Large
0.03
0.45
0.05
Edge, Outgoing
0.08
Core, one direction
14
Example Scenario
—  Small packet has negligible processing delay
—  Small packet experience larger delay at R2 than R1
15
Proportional Energy Savings
—  Reduced delay at R2
More energy savings at R2
—  Increase in multiplexing impact energy savings
—  Relative savings at R2?
Proportional Energy Savings (%)
100
90
80
70
60
1 by 1: R
50
3 by 3: R1
40
1 by 1: R2
30
3 by 3: R2
1
20
10
0 −2
10
−1
0
1
10
10
10
Target Per Router Packet Delay (ms)
2
10
16
40
35
30
1 by 1: Direction−A
3 by 3: Direction−A
2 by 2: Direction−A
25
20
15
10
5
0 −2
10
−1
0
1
10
10
10
Target Per Router Packet Delay (ms)
2
10
Improvement in Energy Savings (%)
Improvement in Energy Savings (%)
Cascading (Domino) Effect
15
2 by 2: Outgoing
3 by 3: Outgoing
1 by 1: Outgoing
10
Rate Scaling: Core
5
0 −2
10
−1
0
1
10
10
10
Target Per Router Packet Delay (ms)
2
10
Active/Idle: Edge
•  Improvement in energy savings
•  Rate Scaling: Up to 35%
•  Active/Idle Toggling: Up to 15%
17
Improvement in Energy Saving
Hybrid Case
12
1 by 1: Outgoing
2 by 2: Outgoing
3 by 3: Outgoing
10
8
6
4
2
0 −2
10
−1
0
1
10
10
10
Target Per Router Packet Delay (ms)
2
10
Hybrid: Edge
•  Improvement of up to 10% observed for hybrid
•  Multiplexing reduces improvement in all three
classes of algorithms
18
Conclusion
—  Performance evaluation framework
—  Trace based analysis
—  Effect of ALR policies on neighboring routers
—  Cascading (domino) energy improvement
—  Up to 30% energy savings (rate scaling)
—  Influenced by traffic characteristics
—  Future Work:
—  Variability
—  Large scale deployment study
—  Interactions with higher layer protocols & applications
19
Thank You
Green Domino Incentives: Impact of Energy-aware
Adaptive Link Rate Policies in Routers
Cyriac James
University of Calgary
Canada
cyriac.james@ucalgary.ca
Niklas Carlsson
Linköping University
Sweden
nikca@ida.liu.se
20