IGCC '14 1
Towards Eco-Friendly Home Networking
Mathias Gibbens, Chris Gniady and Beichuan ZhangDepartment of Computer Science
The University of ArizonaTucson, Arizona
IGCC '14 2
Home networks are complex
● More and more demand is being placed on router performance
● Routers require more computing power, bandwidth and features
NAS
IGCC '14 3
Complexity → Power
Power consumption doubled in 5 years, what about the future?
Always on in 88 million homes, energy footprint of $1 billion
g (2003) n (2009) ac (2013)
Power
CPU
RAM
Dual core!
WiFi Speed
IGCC '14 4
Previous solutions
● Wired: IEEE 802.3az introduced Energy-Efficient Ethernet
▪ First deployed in home networks
▪ Physical connection, easy to detect client
● Large mesh networks: many routers, many clients
▪ Power down redundant access points
▪ Client picks optimal network when more than one is available
● Home networks: one router, many clients
▪ Goma et al: aggregate individual networks
▪ Requires: dense networks, cooperation, client modifications
We need energy management for individual routers
IGCC '14 5
Our contribution
● Transparent energy optimization of personal networks
▪ Individual routers
▪ No user intervention or modification of clients
▪ No cooperation between networks
● Implementation approach
▪ Discarding unnecessary wireless traffic
▪ Powering down routers when idle
▪ Power cycling with active clients
● Increased energy efficiency of individual home routers
IGCC '14 6
Outline
Introduction
Trace collection and categorization
Proposed optimizations
Methodology
Results
Conclusion
IGCC '14 7
Traffic categorization
Traffic seen even without clients connected
Lots of idle time when clients are present
Router is in full power mode independent of clients
T1 T2 T3 T4 T50%
20%
40%
60%
80%
100%
No client broadcast No client Idle client Active client
Tim
e
IGCC '14 8
Eliminating broadcast traffic
Broadcast traffic, but no clients around to respond
Traffic from wired interface retransmitted over wireless
No one listening → drop broadcast traffic
Safe to perform: clients must be present in order to respond
1
10
100
1000
10000
100000
Client Broadcast
IGCC '14 9
Powering down when idle
No clients → power down antennas after a timeout
Periodically check for the arrival of new clients
1
10
100
1000
10000
100000
Client BroadcastLast client departs
Power down
Power up
Power
Look for clients
IGCC '14 10
Optimizing duty cycle: Downtime
● Duty cycle impacts a client's ability to connect
● Need to balance extra delay with potential energy savings
● Infrequent initial associations, which impact only first client
0 1 2 3 4 5 6 7 8 9 10 110
10
20
30
Antenna down time [s]
Co
nne
ctio
n tim
e [
s]
IGCC '14 11
Optimizing duty cycle: Uptime
● Antennas must be up for at least 4 seconds for clients to connect
● From observed delays, we chose a 5/5 second up/down cycle
4 5 6 7 8 9 10 110
10
20
30
Antenna up time [s]
Co
nne
ctio
n tim
e [
s]
IGCC '14 12
Idle connected clients
● During sufficiently long idle periods, turn off transmit antenna
● Possible because either clients initiate or data arrives on wire
● Router must still periodically announce its presence
1
10
100
1000
10000
100000
Client Broadcast
IGCC '14 13
Idle connected clients
● WiFi spectrum is inherently error prone
▪ Existing protocols have built in transparent retransmission to compensate
● Clients typically disconnect from network after 7 seconds
● Transparent retransmission gives us opportunity to power up
Time
ACK
No ACK, try resending Packet lost, inform user
IGCC '14 14
Idle connected clients
● Additional delay due to power up may be seen by applications
▪ Can be hidden in the time it takes for a response to return to the router
● Transitioning router's state has an energy cost
▪ Only sufficiently long periods should be optimized
▪ Ensures real time data is not interrupted
Router power
Client packet
Forward packet
Begin router power up
Complete
Response
Response to client
IGCC '14 15
Methodology
Monitor traffic seen at router: week long traces
Very few initial associations when no other clients present
Detailed power/delay model of ASUS RT-N16 profiled using NI
Trace T1 T2 T3 T4 T5
Average concurrent devices 1 5 4 4 2
Maximum concurrent devices 1 9 7 6 3
Initial associations 13 16 14 31 35
Average time with no client [h] 10.7 0.08 2.07 1.32 0.92
Traffic volume [GB] 5.57 40.21 4.22 3.52 12.97
IGCC '14 16
Eliminating broadcast traffic
No client period increased by 10% average in traces 2-5
More opportunities for router to power down
T1 T2 T3 T4 T50%
20%
40%
60%
80%
100%
No client broadcast No client Idle client Active client
Tim
e
IGCC '14 17
Power cycling no clients
Power cycling with no clients has significant impact
Additional connection delay for first client paid only occasionally
NB PC NB PC NB PC NB PC NB PCT1 T2 T3 T4 T5
0.0
0.1
0.2
0.3
0.4
0.5No client Idle client Active client
Ene
rgy
[MJ]
NB – No BroadcastPC - PowerCycle
IGCC '14 18
Active client optimizations
More aggressive power cycling can reduce energy consumption by an additional 20-30%
Cumulative energy savings observed to be 12-59%
PC
CT
CA
PC
CT
CA
PC
CT
CA
PC
CT
CA
PC
CT
CA
T1 T2 T3 T4 T5
0
0.1
0.2
0.3
0.4
0.5No client Idle client Active client
Ene
rgy
[MJ]
PC – PowerCycleCT – CycleTransmit
CA - CycleAll
IGCC '14 19
Delay due to state transition
Active clients see some delay when initiating activity after a period of idle time
All delays within perception threshold, not noticed by user
0 10 20 30 40 50 60 70 800%
20%
40%
60%
80%
100%T1 T2 T3 T4 T5
Additional delay [ms]
De
lays
se
en
IGCC '14 20
Conclusion
Investigated opportunities to reduce energy consumption of consumer wireless routers
Collected traces from personal networks
Predicted wireless energy consumption reduced by 12-59%
Changes do not break backwards compatibility
IGCC '14 21
Thank You
Questions?
IGCC '14 22
Trace collection and analysis
Five unique week-long traces collected from households
Routers recorded just the wireless traffic seen
Networks used as normal to produce representative traces
1
10
100
1000
10000
IGCC '14 23
Idle periods with clients
To save energy with clients, there must be many idle periods of sufficient length
Each trace has many long idle periods
24
816
3264
128256
5121024
>1024-20%
0%
20%
40%
60%
80%
100%T1 T2 T3 T4 T5
Idle time [s]
We
ight
ed
idle
tim
e