Energy-Efficient Congestion Control Opportunistically reduce link capacity to save energy Lingwen Gan 1, Anwar Walid 2, Steven Low 1 1 Caltech, 2 Bell.
Post on 31-Mar-2015
216 Views
Preview:
Transcript
Energy-Efficient Congestion Control
Opportunistically reduce link capacity to save energy
Lingwen Gan1, Anwar Walid2, Steven Low1
1Caltech, 2Bell Labs
Improve network efficiency by 1000 times!
Network links consumea lot of electricity
Electricity consumptionof network links > Electricity consumption
of the United Kingdom
Fiber optics, copper cable
0%6%
12%
yearly growth rate
Reduce electricity consumption of network links.
Exploit low link utilization
What we do: dynamically manage link capacity.
on average off-peak0%5%
10%15%20%25%30%35%40%
link utilization
Technologies to change link capacity
Link bundle
Sleep mode [Gupta’03]
Voltage and frequency speed scaling [Pillai’01]
component linkto sleep
Link bundle
router router... 2~20
Linear power consumption
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
0.51
1.52
2.53
3.54
4.55
Identical component links
# active component links
Power consumption(units)
energy saving reduced capacity
Outline
• Challenge• Goals• Algorithm• Simulations
Challenge: interaction with TCP
Reduce traffic throughput
TCP reacts
capacity
congestionthroughput
Adjust capacity slowly.
Adjust capacity fast, but TCP friendly.Packet time scale. [Francini’10]…Flow time scale.
• Fast response• Small overhead
Routing time scale. [He’06] [Fisher’10]
Two approaches
This work
Goals
Dynamic Bandwidth Adjustment (DBA) Algorithm, such that
1) Operate at flow time scale.
2) Do not reduce throughput.
3) Save as much energy as possible.
4) Throughput does not oscillate---stability.
Recall TCP
at steady state
transmissionrate
packet loss probability
TCP
packet lossprobability
transmission rate
Recall Random Early Discard (RED)link
incoming traffic link capacity
buffer size
buffer size
packet drop probability
Recall network solves NUM
Transmission rates
Throughput on the links
Ideal throughputIdeal capacity
Thm [Kelly’98, Low’99]: The network model
solves the Network Utility Maximization problem:
Bottleneck & non-bottleneck links
buffer size
packet drop probability
Bottleneck link:
Non-bottleneck link:
buffer size
packet drop probability
• Do not reduce capacity
• Reduce capacity• Keep 0 packet drop
Keep the buffer at the “right” place
buffer size
packet drop probability
targetbuffer
DBA Algorithm (for each link)
1. Pick a target delay satisfying
2. At any time, set target buffer size
current buffer sizecapacity
and update capacity as
zero throughput reduction &maximum energy saving
Thm: Network under DBA algorithm, modeled by
converges to (original) target throughput(zero throughput reduction)
with minimum energy consumption
(maximum energy saving)
Current networkarchitecture
Model network delay
Global stability
TCP sources Links
?
transmission rate
packet loss
incoming traffic
packet drop
No network delay With network delay
delay
Local stability under network delayThm: Network (with DBA)
is locally asymptotically stable,
provided some mild conditions hold.
in the presence of network delay modeled as
Goals
Dynamic Bandwidth Adjustment (DBA) Algorithm, such that
1) Operate at flow time scale.
2) Do not reduce throughput.
3) Save as much energy as possible.
4) Throughput does not oscillate---stability.
✔✔✔✔
Standardsimplifying assumptions
ns2 simulation to verify.
ns2 is a standard and accurate simulation software.
Simulation setup
Node 1
Node 2
1Mb/s
1Mb/s
1Mb/s
1Mb/s
TCP Source 1
TCP Source 20
TCP sink 1
TCP sink 20Compare two configurations• static: 50Mb/s• DBA: 5~50Mb/s
20 additional TCP flows come and go abruptly.
time (s)
Thro
ughp
ut (M
b/s)
staticDBA
Zero throughput reduction
TCP flowscome
TCP flowsgo
Initial dip
Fast recovery
Throughput doesnot oscillate.
throughputpreservation
instantincrease
throughputpreservation
throughputpreservation
capa
city
(Mb/
s)
time (s)
staticDBA
Maximum energy saving
TCP flowscome
TCP flowsgo
capacity rampsup fast
capacity rampsdown slowly
same asthroughput
same as throughput
same as throughput
short transient
Network link is lightly utilized, can reduce capacity to save energy.
Stability: locally asymptotically stable.
Optimality: zero throughput reduction, maximum energy saving.
Concluding remarks
Verified by ns2 simulations.
Propose DBA to adjust link capacity in TCP flow time scale.
Thank you!
top related