Energy-Efficient Design Some design issues in each protocol layer Design options for each layer in...

Preview:

Citation preview

Energy-Efficient Design

• Some design issues in each protocol layer

• Design options for each layer in the protocol stack

Energy Consumption Here are some results on energy consumption

of networking hardware:

Device Sleep power (mW) Idle/Wakeup Power (mW) wakeup time (ms)

WaveLAN 143 1148.6 100

Metricom 93.5 346.9/431 5000

IBM IR - 349.6 100

Newton PDA 164.2 1187.8 N/A

Magic link

PDA 312 700 N/A

Laptop 8000

Energy Consumption

• WaveLAN: transmission power : reception power = 1.4 : 1

• Rockwell sensors: transmission power : reception power = 4:3 (full range); 1:1 (min range)

• Berkeley’s motes: operate at around 10mW.

Mechanisms to reduce power at physical layer

• First-order approximation of the power consumption of CMOS circuitry:– P = C * V^2 * f (C: effective switch capacitance; V is the

supply voltage; f is the clock frequency)

• Ways to reduce power:– Reduce the supply voltage V; (cons: may reduce performance

at the same time, need perf. compensation)– Reduce the switching frequency f (sleep/idle mode); (cons:

no energy reduction due to longer runtime)– Reduce the capacitive load C

• Reduce external access (output, memory, etc.)• Reduce logic state transitions, routing capacitance

Reducing power at MA C and link layer

• Principles:– Only power on when transmitting– Increase the ratio of effective transmissions and avoid

retransmissions, collisions, corruptions– regulate traffic

• Mechanisms to achieve it– Add idle/sleep mode in the transmission state diagram– Regularize transmissions and make them predictable, nodes

can sleep when not transmitting– Avoidance contention loss, congestion loss, retransmissions,

error-prone transmissions, as much as possible• how to achieve ??

Reducing Power at Network Layer• Principles:

– Put routers that do not forward packets into sleeping mode• SPAN, GAF, PEAS

– Reduce signaling packets (e.g. LSP)– On-demand routing– Packet size and fragmentation– Energy-related optimality criteria– System decomposition: do not stress energy-critical nodes, load

balancing– Use intermediate nodes as middleman

• Key issues:– How to guarantee connectivity

PEAS: Probing Environment, Adaptive Sleeping

• Enable long-lived system by exploiting the scale

• How?

– Keep a subset working, turn off the others into sleeping mode

– Sleeping ones replace dead ones

• Two issues

– working nodes should be evenly distributed across the field: Probing Environment

– Replace failed or energy-exhausted working nodes quickly: Adaptive Sleeping

Distribute working nodes evenly: Probing Environment

• Nodes are in sleeping initially– An exponential random time

• When waking up, a node probes within a range R_p– If there is no working node

within R_p, it starts working

– Otherwise, the working one should send back a reply and this one sleeps again

Peas model: why working node are evenly distributed

• Any topology of working nodes is equivalent to placing round peas of radius R_p/2 on a plane

– The centers of two peas are at least R_p apart, when they’re tangent

• Assuming infinite deployment density

– The densest case: each pea is tangent to 6 neighboring peas

– the sparsest case: the space among any 3 adjacent peas is slightly smaller to insert another pea

Detect unpredictable failures: randomized sleeping times

• If wakeups are synchronized to some time points– long “gaps” if a working

one fails unexpectedly

• Randomized sleeping times spread wakeups over time– Any unexpected failure is

detected on time

How to adjust the wakeup rates• A working node

measures the aggregate wakeup rate from its probing neighbors and

• includes the information in the replies to probing neighbors

• Each probing neighbor adjusts its rate accordingly

TsTime

K wakeups

Measure aggregate rate: _a = K / (t - t0)

t0 t

Each probing one adjusts: _new = (_d / _a )

Reducing power at the transport layer

• Principles:– Avoid packet loss and retransmissions

– Reduced-power operations during backoffs

– Light-weight connection management

– Design balance at the sender and the receiver

• Mechanisms– State estimation: Do not transmit during channel errors

– Congestion avoidance: Do not overload the network

– Buffering

Reducing Power in the OS (including the FS)

• Principles:– Predictable schedules– Caching– Turn off idle devices (e.g. displays)– CPU scheduling via dynamic voltage scaling

• Mechanisms:– CPU scheduler: idle state– Use application hints & semantics

Recommended