10-white-lacnog-ener.. - LACNIC

thoughtlessskytopΔίκτυα και Επικοινωνίες

29 Οκτ 2013 (πριν από 3 χρόνια και 7 μήνες)

55 εμφανίσεις

ENERGY
AWARE
ROUTING

Russ White

russw@riw.us

Motivation


Optimal network
performance


Optimal resource use


Drive the load adaptive
principles of
EnergyWise into the
network

Load
Adaptive
Networking

ADAPT

CONSERVE

EDGE

NETWORK

Motivation


The initial round of
energy savings is taking
place at the edge


But what about the
core?


Network elements and
links are expensive to run
(power wise)


The problems here are more
complex to resolve


But the savings could still be
significant


Highest Cost

Intermediate Cost

Lowest Cost

Motivation


What are the power
savings possible by
reducing usage in the
core?


These numbers from a
Cisco 12000 give us
an idea of the scope
and level

Power Management for Networks to Reduce Energy
Consumption, David
Wetherall

Motivation


Energy cost is a factor
of capacity, not use


Matching capacity
closely to demand
saves energy


In other words, adapting
the network capacity to
the network load

Energy Consumption (Watts)

Maximum Throughput (Mb/sec)

GreenComm
-
ICC09
-
Keynote2
-
Nordman.pdf

Motivation


The techniques
presented here can
reduce energy usage
by about 20% in a
typical campus


$350 per year per
network device


Conservative estimate


But our competitors
agree with us…

Network is designed for this level of traffic

Load Adaptation could save these resources

Existing Approaches
MANET


Minimize per packet
costs


Generally used in
MANET networks


Use nodes based on
battery power


Why isn’t this useful for
wired networks?

Avoid this node

Prefer these nodes

Existing Approaches
MANET


Different network conditions....


MANET devices are battery powered


Wired devices are “plugged in”


Lead to different goals


MANET


Extend battery life


Per node power usage is important


Spread usage among many nodes as much as possible


Wired


Reduce overall network usage

Load Adaptive Control Plane


When does a router stop
switching traffic?


In a modern network design

never!


Load sharing means most links are
used all the time


When could a router be placed
into sleep mode?


In a modern network design

never!


The goal of the load adaptive
control plane is to minimize the
network topology


Dynamically


Based on business requirements
specified by the user


Based on dynamic load information

Load Sharing

Load Adaptive Control Plane


Reduce the topology to the minimal
required for the offered/expected
load


Expected load based on user built
profiles


Time of day, resiliency requirements,
etc.


Offered load could be based on real
time traffic measurements


The control plane can determine the
minimal topology required


Routing, such as EIGRP, OSPF, etc.


Switching, such as TRILL, Spanning
Tree, etc.


The control plane does not “put
devices to sleep”


Simply builds the opportunities for
devices to use built in energy efficient
modes

Minimal Topology

Load Adaptive Control Plane

Minimum
Topology

Secondary
Topology

Full Topology

Notes

Fully

Enabled

Fully Enabled

Fully Enabled


No recovery time


No

reduction in bandwidth


Survives

two failures

Fully Enabled

Fully

Enabled

Low Energy Use


Minimal

recovery time (<1 sec)


Small reduction in bandwidth


Survives single failure

Fully Enabled

Low

Energy Use

Minimum

Energy
Use (“Off”)


Moderate recovery

time (1
-
3 sec)


Moderate

reduction in bandwidth


Failure requires recovery

Low Energy Use

Minimum Energy
Use (“Off”)

Minimum Energy
Use (“Off))


Long

recovery time (minutes)


Minimum

bandwidth available


Failure requires recovery

Weekend network posture

“After Hours” network posture

Normal Network Posture

Energy Managed Networks


Challenges


How do we build these “energy states?”


How do we build these “network topologies?”


How do we make all of this work together?

Building Energy States


Power levels defined
through standards
process


H is High


R is Reduced


F is Frugal


These are just examples


Power states
correspond to ideal
network states


Edge Devices

Aggregation

Core

Normal Posture

F

R

H

Recovery Time

0 seconds

Off Hours Posture

F

F

R

Recovery Time

< 1

second

Weekend Posture

F

F

R

Recovery

Time

<

2 minutes

Building Energy States


The network manager sets various “state triggers”


Done through network management


For instance, time of day, etc.


When a trigger engages


The network management station sets the devices within each
“zone” to the appropriate “ideal state”


The control plane then works to move the network into the
ideal state


Examines actual network usage, etc.

Building Energy Topologies


The process of moving a network into a desired state
relies on being able to find reduced energy network
topologies


To move the network into a given state, the devices in a given
topology are moved into the appropriate state


Three mechanisms


Traffic Engineering


Routing Modifications


Disjoint Topologies

Building Energy Topologies
TE


Use traffic engineering
to push traffic onto a
subset of links in the
network


Openflow

or MPLS/TE


Links and devices not
used can be placed
into sleep mode

Building Energy Topologies
RP


Modify the existing routing protocols to find a “minimal
set” of the topology


The amount of redundancy, bandwidth, and other factors
can be controlled dynamically

Building Energy Topologies
Disjoint


This is a relatively new idea in the network world


Use an algorithm to find two topologies with non
-
overlapping links
within the network


Mark each topology so the routers treat each one as a logical
network in some way


Route over each topology independently


Still very “researchy”


Some algorithms in this space wouldn’t work well for our purposes


Something to keep an eye on and think about, rather than take
action on “right now”

Bringing Devices Back


Once a device is asleep, how do we bring it back up?


Two general mechanisms


Out of band signaling


SNMP,
Openflow

with an “alive” RP,
etc


In band signaling


Wake on LAN capability built into the box


Either way, this piece needs hardware modifications


Real functionality in this space won’t exist until vendors are
convinced there enough cost/benefit ratio to justify the hardware
changes

Summary


There are significant savings possible


Savings at the edge swamps the network savings, however


It’s going to take lots of work to get to the savings in the core


Network management controls the ideal state


Signals the control plane it’s okay to drop to a lower power state


Signals the control plane when it should move to a higher power
state


The control plane controls the actual state


Within the parameters given by the ideal state


A “range of states” may be possible, as well