Upload
vuque
View
215
Download
2
Embed Size (px)
Citation preview
Power Management Options forMobile Wireless Devices
Matthew S. Bosworth
2000
Advisor: Prof. Siewiorek
,~t~ Electrical & ComputerENGINEERING
Power Management Options for Mobile Wireless Devices
Matthew Bosworth
December 10, 1999
Abstract
With the proliferation of devices such as mobile phones, pagers, palm top computers, personal
digital assistants (PDAs), and now the introduction of wearable computing devices, an important
question is "how are these devices to be powered?’ Currently it is common for over 50% of the
weight of these devices to be power related[13]. Further, a main concern of users and designers of
these systems is effective battery life. This research investigates two methods to reduce the power
consumed by (and thus increase the battery life of) these devices.
The first method explores the inclusion of a "super-capacitor’ in the battery pack. Recent re-
search in non-ideal battery modeling shows that, under certain circumstances, computations per
battery life has a closer relation to peak power than to average power. This paper shows that the
technique of adding a super-capacitor can reduce high power spikes, increasing effective battery
life by up to 10%.
The second method examines the wireless networking capabilities of current mobile systems,
and discusses how these capabilities affect battery life. It is shown that while wireless commu-
nications can be a battery drain, there is a space wherein it can be used to increase computations
per battery life. This increase comes through remote processing of jobs. The boundaries of this
space are explored and discussed, and a basic equation for deciding whether or not to process a
job remotely is derived. Using the measured characteristics of an available system, the equation
shows that a job must process for at least 1.4 milliseconds for each kilobit transferred wirelessly in
order for remote computation to reduce overall energy consumption.
Contents
1 Overview 4
1.1 Introduction .......................................... 4
1.2 Experimental Setup ...................................... 6
2 Hardware Power Management 8
2.1 Batteries : A Brief Review .................................. 8
2.2 Battery Requirements for Mobile Computing ....................... 10
2.3 E~ergy Characteristics of Secondary Batteries ....................... 11
2.4 Load Leveling with Super Capacitors ........................... 13
2.5 Measurements ......................................... 14
2.6 Results and Conclusions ................................... 18
3 Power Management via Remote Processing 21
3.1 Wireless Communication: A Brief Review ......................... 21
3.2 Computation versus Communication ........................... 23
3.2.1 Power Consumption ................................. 26
3.2.2 Latency ......................................... 27
3.2.3 Network Conditions ................................. 28
3.2.4 Comparative Processing Power ........................... 29
3.2.5 Threshold Calculation ................................ 29
3.3 Model Results ......................................... 20
3.4 Results and Conclusions ................................... 32
34
Appendices 35
A Matlab code for Local/Remote Algorithm 36
B Glossary 37
2
List of Figures
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8
2.9
2.10
li-ion Voltage versus Time for a Constant Current Load ................. 9
li-ion Voltage and Current versus Time for Constant Power Load ........... 11
Capacity versus average power for a standard battery pack ............... 12
Block diagram of the Spice simulation circuit ....................... 14
Capacity versus average power for various hybrid battery packs ............ 15
Power drawn by various hybrid packs at 10% duty cycle ................ 16
Power drawn by various hybrid packs at 80% duty cycle ................ 17
Battery life versus duty cycle for various hybrid battery packs ............. 18
Battery life improvement for a 20F hybrid pack ...................... 19
Mass and volume versus capacity for hybrid packs .................... 20
3.1 Data set size versus computation time thresholds ..................... 31
Section I
Overview
1.1 Introduction
According to Dr. Dan Stancil of Carnegie Mellon University, in the not too distant future
people will look back at the tangled mess of their wired lives and laugh. Yet, in the meanwhile
all of the current wireless devices are the victims of insufficient battery life [16]. A report by the
National Research Council notes that not only do batteries make up over half the weight of many
wearable wireless computing devices, but after a few years the cost of replacing the batteries in
these devices can exceed the original cost of the unit! [13]
Statements such as these make the motivation for this project very clear. As the world moves
away from wires, the need for high-capacity mobile power systems increases. Dr. Theodore Rap-
paport claims that wireless is "enjoying its fastest growth period in history"[ll]. Furthermore,
while the ratio of power consumption to millions of instructions per second (MIPS) is falling
a factor of ten every five years, there are fundamental limitations on radio technology which will
prevent a similar improvement for wireless transceivers. Thus future generations of wearable and
mobile devices which communicate wirelessly will expend an increasingly larger percentage of
4
their battery power in order to do so. It has been estimated that wireless communication will be
responsible for 80% of mobile power consumption by 2001 [13]. In order to reduce these and other
problems caused by increased reliance on batteries, several avenues are being explored.
The obvious approach to improving batteries is chemistry research. Unfortunately, this avenue
seems to be approaching a point of diminishing returns. In fact, battery capacity has only doubled
in the past 40 years [13]. One of the main reasons for this seems to be that the amount of energy
stored per kilogram of li-ion battery is quickly approaching the energy stored in the same mass
of explosives [13]. Packing more energy into the same space has become a very difficult problem.
Also, the problems of safe handling and disposal also bear serious consideration.
Another possibility for the extension of battery life is low power electronics, as can be seen
by the dropping voltages of processors and other support chips. For some chip families, in each
successive generation clock speed improves in combination with reduced power consumption
[12]. Based on this a new, energy-based Moore’s Law has emerged. From 1990 to 1995 the power
to performance ratio of chipsets has fallen by an order of magnitude every 30 months, and is
expected to drop by another order of magnitude between 1996 and 2001 [13].
In this paper, two other avenues will be explored. First, recent work in non-ideal battery
modeling has shown that, in some cases, battery life is more dependent on peak power than on
average power. To date, most estimates of battery life have been made based on average power
requirements which can over-rate battery life by as much as 50% [9]. Based on the premise that
peak power is the determining factor in battery life, this project will look into ways to reduce the
peak power of a system as seen by the battery.
Recent breakthroughs in material science have brought forth a new type of capacitor of hereto-
fore unheard of energy density. Such a capacitor in parallel with a battery, absorbs some of the
energy of a power spike and spreads it over time. Thus, the average power increases, while peak
power drops. As will be seen shortlj~ battery power often correlates to peak power, so these ca-
pacitors can effect an increase in effective battery capacity.
Second, the ability of mobile platforms to communicate wirelessly with server systems will
be explored in a power management context. If a small data set of an intensive algorithm is
sent to a system which is not subject to power limitations, battery power can be conserved. This
research presents an equation to determine whether a given task would extend battery life if it
were computed remotely.
1.2 Experimental Setup
The mobile device which all of the simulations and measurements in this report have been based
on is the Compaq Itsy computer. The Itsy has a clock-throttleable StrongARM processor with
eight megabytes (MB) of flash memor~4 64 MB of random access memory (RAM), three serial
ports, and a gray-scale touch screen display. At the highest clock frequency the processor is rated
at approximately 90 Dhrystone MIPS. The Itsy runs the Linux operating system, weighs about 120
grams, and consumes less than one watt.
The Wearables Lab at Carnegie Mellon University has added a PCMCIA daughtercard to the
basic Itsy design to enable wireless access via a WaveLAN PCMCIA card. The WaveLAN card
works in three modes : sleep, transmit, and receive. However, the Linux driver which was used
to test the power consumption of the card does not implement the sleep mode, so the empirical
results presented here are based entirely on the latter two modes. The implications of this are
discussed in Section 3.2.
In addition to the experimental apparatus, a Spice simulation was used to model the behavior
of batteries and capacitors which were not available. The battery simulations have two main
components, the battery pack, and the load. The Spice model used to represent these is discussed
in Section 2.4. The simulation was performed with Avant!’s HSPICE simulator on a cluster of Sun
SPARC machines.
Section 2
Hardware Power Management
2.1 Batteries : A Brief Review
Many volumes have been written on the various properties of both primary and secondary bat-
teries. For the purpose of this paper, a few working definitions have been presented in Appendix
B. Some of the pertinent and high level properties will be discussed in this section.
The battery properties which are of greatest concern to mobile systems designers as well as
users are capacity and voltage. Weight, volume, and safety (of use and of disposal) also bear
consideration [13]. A battery’s capacity can refer to either charge capacity or energy capacitN
which are similar but distinct quantities. Charge capacity is measured in Ampere-hours (Ah) and
indicates the length of time a battery can be discharged at a specific current. Charge capacity is
also referred to as a battery’s "C rating." Energy capacity is the charge capacity multiplied by the
nominal voltage of the battery, and thus is measured in watt-hours (Wh). A battery’s voltage
the potential measured between the terminals of the battery~ and is fairly consistent for batteries
of a particular chemistry. However, a battery’s voltage is not constant over a discharge cycle. This
can be seen in Figure 2.1. Here, the voltage falls off non-linearly over the course of a constant
current discharge. Once a battery’s voltage falls below a specified value, 2.4 volts in this case, it is
considered discharged.
4.2
3,8
3,6
3.2
3
2.8
2.6
2.40 3500500 1000 1500 2000 2500 3000
Time (seconds)
Figure 2.1: Voltage versus Time for a li-ion Battery with a Constant Current Load. Note that thevoltage scale does not begin at zero.
Other properties, such as weight and volume, are usually conveyed in reference to energy or
power, since batteries are available in a wide variety of form factors. The four most common
statistics are specific energy (Wh/kg), energy density (Wh/L), specific power (W/kg), and
density (W/L)1. Lastly, safety is one of the most overlooked battery traits. The heavy metals, such
as lithium and cadmium, which are used to create many re-chargeable batteries are not easily
disposed of once the useful life of the battery is oven Furthermore, about the only difference
between a pound of the latest lithium ion batteries and a pound of dynamite is the rate of energy
watts, Wh = watt.hours, kg = kilograms, and L = liters
discharge - the specific energies are comparable [9]. That said, safety and disposability will not be
discussed further.
A battery’s voltage is not constant over a discharge cycle. This can be seen in Figure 2.1. Here,
the voltage falls off non-linearly over the course of a constant current discharge. Once a battery’s
voltage falls below a specified value, 2.4 volts in this case, it is considered discharged.
2.2 Battery Requirements for Mobile Computing
Mobile electronic devices, especially those which interact directly with people, do not present a
constant current load. To the contrary, mobile electronics generally use a switching power supply
which looks (to the battery) like a constant power load. As can be seen in Figure 2.2, as the battery
voltage drops an increase in current is required to maintain constant power. However, as the user
places different demands on the electronics, the "constant" power seen by the battery fluctuates.
To elaborate, consider the average laptop. The laptop is composed of several subsystems, each
of which only has to be on (consuming power) when it is needed. These subsystems include
hard disks, CD-ROM drives, floppy disk drives, displays, and PCMCIA cards. Much of the time,
several or even all of these subsystems are partially or completely powered down. When all of the
subsystems (including the CPU) are in their minimum power states, the power drawn from the
battery is at a minimum - this state will be referred to as ’idle." When demands are made on the
system requiring various subsystems to power on, the power drawn from the battery can fluctuate
rapidly and significantly. For example, the Compaq Itsy has an idle power just under half a watt.
This power draw triples when the Itsy is decoding MPEG-2 video. Note also that the Itsy is not
equipped with a spinning disk drive nor a color display, which make up a considerable part of the
power draw on most laptop computers [12].
10
0.9
0.8
0.7
0.5
0.4
0.3
0.2
0.1
4.510~00 20’00 30~00 40~00 50~00 60~00 70~00 8000
4
2.5
2 ~ 40~00 60~000 1000 20~00 30~00 50~00
Time (seconds)70~00 8000
Figure 2.2: Current Vs. Time (above) and Voltage versus Time (below) for a li-ion Battery with Constant Power Load. Note that the voltage and current scales do not begin at zero.
2.3 Energy Characteristics of Secondary Batteries
Figure 2.3 illustrates how a pulsed power load can lessen the energy capacity of a standard sec-
ondary battery. The solid line depicts the delivered energy capacities (Wh) of a li-ion battery
being discharged at various constant power loads. Note that as the load power increases, the total
energy delivered by the battery decreases. The z~’s show the same battery under a pulsed load
characterized by a 450 milliwatt minimum power and 10 watt maximum power (from left to right,
the/~’s indicate 10%, 50%, and 80% duty cycles for the load). Note that the energy capacity for
pulsed load (e.g. the 10 watt peak load with 10% duty cycle represented at point A) is much less
than the energy capacity for a constant load with the same average power (e.g. 2 watt constant
power at point B). Moreover, note that for all three duty cycles, the pulsed power capacity is close
to the capacity of a battery undergoing constant power discharge at the maximum power of the
11
pulse (e.g. point C for the 10 watt peak discharge). A similar phenomenon (though less noticeable)
occurs at the ’+’ marks, which indicate a pulsed discharge with an idle power of 450 milliwatts
and a peak power of 1 watt. It is also noteworthy that, if the duty cycle and pulse power are of
sufficient magnitude, the battery capacity driving a pulsed load is actually less than the capacity
of the battery driving constant load of magnitude equal to the peak pulse power. This is due to
several non-ideal battery properties including concentration polarization, ohmic polarization, and
activation polarization. These properties are discussed in detail by Dr. Martin in [9].
3.4- ~ ~ ~ t I0 2 4 6 8 10 12 14 16 18
Average Load Power (watts)
Figure 2.3: Capacity versus average power for a standard battery pack under constant power andpulsed power loads. Note that the battery capacity scale does not start at zero.
12
2.4 Load Leveling with Super Capacitors
Based on the information presented in Section 2.3, it was theorized that large capacitative de-
vices could reduce the peaks seen by the battery pack, and thereby allow a pulsed power load to
approach the capacity of a constant power load. Recent advances in capacitor technology have
allowed large capacitors to come in small enough form factors that they are usable in mobile wire-
less devices [7], [8].
The super capacitors now available have capacitances of up to 20F and are slightly larger than
a AA battery. A capacitor of this size is able to reduce or remove short duration power spikes,
at the cost of raising average power slightly. Unfortunately, limited production has reduced the
availability of these capacitors. For this reason, the data presented herein comes from simulations.
Using a Spice model for the Li-ion battery, several different hybrid power supplies were modeled
at various duty cycles and peak power requirements. A block diagram of the simulated circuit
is shown in Figure 2.4. The left hand side is a model for a constant power load taken from the
work of Dr. Tom Martin [9]. From left to right, the independant voltage source is piecewise linear
(PWL) and is used to set the load power. The load power has three important characteristics: the
minimum power, the maximum power, and the duty cycle. The maximum power and the duty
cycle were adjusted to simulate different loads. The minimum power was set at 450 milliwatts for
all the tests, since this is the measured idle power of the Itsy. The second source is a dependant
voltage, equal to the power drain of the load divided by the voltage across the terminals of the
battery. The rightmost source sets the current across the battery equal to the voltage of the second
source. Since current can be defined as power divided by voltage, this circuit models a constant
power load by increasing the current in direct relation to the falling battery voltage. The right hand
side of the diagram is the hybrid battery module, simply a battery in parallel with a capacitor. Note
13
that the battery here represents a complex model based on the ICR-18650 li-ion battery. The model
was developed by PolyStor Corp. and is presented in [5]. The battery itself has an estimated
series resistance (ESR) of .08f~ and a capacity of 1.25 ampere-hours (Ah). The capacitor half of
battery pack represents the independent variable in the trials. The capacitor was modeled with
values of 1, 10 and 20 farads.
Constant Power Load Hybrid Battery
Pack
, ,PWL. ~ y V(IO) V(20) I II -’~-<xqY) ,,,,
0 ,, 0
Figure 2.4: Block diagram of the Spice simulation circuit
2.5 Measurements
Figure 2.5 shows the progressive improvement as larger capacitances are used in the hybrid pack.
Note that as the capacitances increase, the pulsed load capacities begin to approach the capacity
of the constant power load at the same average power. It is also noteworthy that the gains do not
diminish with larger capacitances. In fact, it can be seen that the one farad capacitor makes little
or no difference in effective battery life. Larger capacitors weren’t measured, mostly because 20
farads is the largest capacitor currently available in a form factor useful for mobile devices.
Figure 2.6 shows the effect that the super capacitor has on the power seen by the battery in
the time domain. Note that the curve without the capacitor has a short duration and high peak
power, whereas the curves with the large capacitors have longer durations and lower peak power.
Essentially, the capacitor has spread the power pulse over time. The area under each of the curves
is roughly the same, but the height of the peak has been reduced. Based on the battery properties
14
Constant Power Load-~ ....... .~- 20F Hybrid BatteryO ....... © 10F Hybrid Battery
............................................ ~ ...... ~ 1F Hybrid Battery4.6 : : : : : ~. ...... ~k Standard Battery
~4.4
g4.~
3.8
3.6
3.42 4 6 10 12
Average Load Power (watts)14 16 18
Figure 2.5: Capacity versus average power for various hybrid battery packs under constant powerand pulsed power loads. Note that the battery capacity scale does not begin at zero.
presented in Section 2.1, this lower peak power explains the capacity improvements seen in Figure
2.5.
Figure 2.7 is similar to Figure 2.6, with a higher duty cycle. Because of the higher duty cycle,
the 10F capacitor almost completely discharges during each power spike. In the case of the 1F, it
discharges several times over. As a result, in the case of the one farad capacitor, the battery has to
cope with both the peak load and recharging the capacitor simultaneously, making it a liability in
high duty cycle applications.
Figure 2.8 displays the improvement in battery life for different capacitors at low power re-
quirements. Notice that the five lines are almost indistinguishable at each duty cycle.
Figure 2.9 shows the improvement that a 20F hybrid battery pack gives over a standard battery
pack. Note that a 15 watt peak at a 10% duty cycle only yields a 10% battery life improvement.
15
0.9
0.8
0.6
0.5
/7"] I -- No Capacitorl/I I -’ 1 Farad Capacitor
i,I - - lO_Farad Capacitor
890 895 900 905 910 915Time (seconds)
Figure 2.6: Power drawn by the battery in parallel with various super capacitors with a constantpower load of 10% duty cycle and I watt peak.
This maximum improvement occurs at a 10% duty cycle (and not at 1%, as might be expected),
because lower duty cycles begin to fall out of the range where capacity correlates more to peak
power. According to Dr. Martin, capacity correlates to peak power at duty cycles of less than 1
cycle per second, and at loads greater than 0.1 C. Remember that 0.1 C is a current of greater than
one-tenth the capacity (in ampere hours) of the battery [9]. In this case, the battery has a capacity
of 1.25 ampere hours, and a 10% duty cycle load presents an average current of
((1W ̄ .01) + (.45W ̄ .99))/3.5V
Since .11 amperes is less than .1 C (.125 amperes for this battery), a 10% duty cycle will not correlate
to peak power.
Figure 2.10 shows possibly the most important tradeoffs; weight versus battery life, and vol-
16
1.02
1
0.98
~-~ 0,96
0.94
0.92
0.9
3220~
No Capacitor-~- 1 Farad Capacitor- ¯ 10 Farad Capacitor
’ 3230’ 32~353225 3240 3250 3255 3260 3265Time (seconds)
Figure 2.7: Power drawn by the battery in parallel with various super capacitors with a load of80% duty cycle and I watt peak. Note that the power scale does not begin at zero. This graph is amagnification of the peaks, the minimum power draw is not pictured.
ume vs. battery life. The numbers in this chart are for a 10 W peak load at 10% duty cycle. As
noted earlier, a 10% duty cycle is near ideal, and a 10 W peak, while high, is not completely un-
reasonable. The peak in the middle of the chart is caused by the four farad capacitor, which is
only available in a prismatic, rather than cylindrical, enclosure. The prismatic cases have a larger
volume, but fit better into most electronic devices, being rectangular. Further, the prismatic enclo-
sures are steel rather than aluminum, which explains the weight difference.
Based on Figure 2.10 it seems unlikely that mobile device designers would use a super capac-
itor in a device, since the addition of more batteries would always be more advantageous. Based
on the same load as Figure 2.10, Table 2.1 shows the improvement necessary in density (i.e. Farads
per Liter) necessary for a super capacitor to become a viable option. That is, for a fixed volume of
batteries, if a one Farad capacitor can improve from its current density (approximately .58 Farads
17
x 104
.~3.5 .................................................................... ~
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9Duty Cycle
Figure 2.8: Battery life versus duty cycle for various hybrid battery packs with a load of I wattpeak power. Note that the battery life scale does not begin at zero.
per 50 milliliters) by a factor of 250, then it would provide the same energy boost as an equal
volume of batteries.
2.6 Results and Conclusions
From a theoretical perspective, the original hypothesis has been upheld; the addition of a super
capacitor to a battery pack lowered the peak power seen by the battery and resulted in an increase
in energy delivered by the battery. With a large enough capacitor, the capacity of the battery
begins to look like a constant power load of average power, rather than a constant power load of
peak power. However, from an engineering perspective, the number of applications that can take
advantage of a super capacitor is limited. The largest capacitor at the highest peak power and
ideal duty cycle (i.e. the most effective arrangement) provides only a 10% improvement in battery
18
0,12-e- 1W Peak--x-- 2W Peak~ 3W Peak~ 5W Peak--8- 10W Peak-e- 15W Peak
0.08
0.06
0.04
0.02
Duty Cycle
Figure 2.9: Improvement in battery life for a 20F hybrid battery pack over a standard batterypack at various loads. The x axis represents the different duty cycles, while each line represents adifferent peak load power.
life, approximately 18 extra minutes in 3 hours. This minor improvement comes at the cost of a
larger form factor, greater weight, and greater cost for the system as a whole. Furthermore, until
capacitor density improves by approximately two orders of magnitude, the volume occupied by
a capacitor would be better used for an equal volume of batteries.
19
~ 85
~ 60
45
2.8 2.82 2.84 2.86 2.88 2.9 2.92 2.94
,~ 26
~ 26~8
2.8 2.82 2.84 2.86 2.88 2.9 2.92 2.94Battery Pack Capacity (hours)
Figure 2.10: The upper chart shows mass versus battery life for a 10 W peak load at 10% dutycycle. The lower chart shows volume versus battery life for the same load. The anomaly at 2.84hours is caused by a particular capacitor which is prismatic (rather than cylindrical) and enclosedin steel (rather than aluminum).
Current F/50mL Required F/50mL Percent Improvement Required1 Farad 0.57803 144.014 249.14%1.5 Farad 0.63559 151.665 238.62%4 Farad 0.34423 186.018 540.38%10 Farad 1.29807 253.577 195.25%20 Farad 1.20919 358.782 296.71%
Table 2.1: Current and required capacitance densities. The first column lists the capacitance per 50milliliters of current super capacitors. The second column lists the density required for a capacitorto be as useful as an equal volume of batteries in a fixed volume battery pack. The third columnshows the percent improvement requi.red for today’s capacitors to become viable in a fixed volumebattery pack
20
Section 3
Power Management via Remote
Processing
3.1 Wireless Communication: A Brief Review
Wireless communication, in various forms, has been around since the late 1800’s [11]. While it is
well beyond the scope of this report to go into a detailed history, below is a short description of
the three types of wireless which are currently most common in the context of mobile computing.
Cellular radio, with its popularity for telephony, is now an option for mobile data communi-
cations as well. Cellular has two main advantages over the other forms of wireless. First, it is
extremely long range relative to its competitors. Also it is widely available, with cells covering
most of the United States and spreading quickly. However, there are some serious drawbacks as
well, not the least of which are the significant costs associated with service and connection time.
While costs are dropping, there still remain other constraints. Bandwidth is limited. The medium
is noisy and thus prone to data errors and the need to resend information. Lastly, the most com-
mon protocol over voice cellular networks is Cellular Digital Packet Data (CDPD). CDPD is
21
second priority protocol on most networks. That is, the CDPD radio is only allotted a channel
when there is one unused by a voice call, and will be cut off when a new call is in need of that
channel [11].
Infrared is another option available for sending and receiving data wirelessly. It is very low
bandwidth, short range, and line-of-sight. In fact, the only significant advantage that infrared has
is its low power requirement. It is mentioned here only for completeness.
The third transmission method, spread spectrum radio, is the main concern of this report.
IEEE 802.11, or Wireless Ethernet, is a fairly new standard for medium range, high bandwidth,
spread spectrum data communication. The standard defines protocols for three different physical
layers: direct sequence spread spectrum (DSSS), frequency hopping spread spectrum (FHSS),
infrared. Several companies have built solutions which meet this specification, but this paper will
concentrate on one configuration : Lucent’s WaveLAN PCMCIA card for the mobile client, and
Wireless Andrew, the Carnegie Mellon wireless network (which consists of WaveLAN wireless
access points) as the server side. Lucent has chosen to support DSSS, as opposed to FHSS, for the
physical layer.
WaveLAN supports a link of up to two megabits per second (Mbit/s) with a maximum range
of 200 meters. The IEEE media access control (MAC) protocol for 802.11 is discussed in [3]. The
capabilities of the WaveLAN card and access point are presented in [1] and [2].
The IEEE standard also defines three operating modes for clients: sleep, transmit, and receive.
Sleep mode is defined specifically for power management. When the client announces that it will
be dozing, the access point agrees to buffer all of the client’s packets. The client and the access
point then agree on an interval at which the client will wake up and listen for a beacon from the
access point. The beacon will broadcast which clients have buffered packets. This protocol allows
the card to power down everything but a timing circuit, creating a significant power savings [3],
22
[17J. For example, the WaveLAN card is rated at nine milliamps at five volts in sleep mode (as
opposed to approximately 300 milliamps in the other two modes). Transmit mode is used to send
data at a rate of up to ~two Mbits/s. The receive mode is used to receive packets from a remote
host at the same rate. The difference between the modes (besides the direction of data flow)
the power drawn by the card. Transmit mode requires approximately 330 milliamps at five volts.
Since reception is a more passive process, the power requirement drops to approximately 280
milliamps at five volts. Note that packet reception is not a completely passive process, since the
IEEE 802.11 standard requires an ACK (acknowledgement) packet be sent after each successfully
received packet.
3.2 Computation versus Communication
Throughout the history of computing devices, there has been a so-called ’wheel of reincarnation’
[15]. In the 1960’s dumb terminals were used to run jobs on mainframes. In the 1970"s mini-
computers were developed to provide local computation and offioad the mainframes. In the late
1970’s and early 1980’s terminals were attached to minicomputers. The circle is occuring now with
handheld PDAs and mobile thin clients.
Thin clients, by their very nature, are one end of the computation versus communication spec-
trum. Containing no computing power of their own, they transmit most or all of the data they
collect to a remote server, usually wirelessly. An extreme example of this is the InfoPad project at
UC Berkeley. The InfoPad includes a touch screen display, microphone, speakers, and a wireless
antenna. The only computing done on the device is driving and sensing the various I/O devices.
Each pen stroke on the touch screen is sent to a remote server for interpretation. The server then
replies with data to be displayed on the screen (i.e. the pixels under the pen should be drawn in
23
black) [10]. Other examples of this sort of device are low-end cell phones, one way pagers, and
global positioning systems (GPSs).
The other end of the spectrum are stand-alone hand-held devices, such as the first iterations
of the Palm Pilot. Most of these devices have no way to communicate with other computers
unless wired in (usually via serial port). However, they are equipped with computational engines
sufficient to handle any tasks required of them.
The latest generation of mobile wireless devices, which are the concern of this report, contain
both wireless communication capability and computational power. While this combination seems
to offer the best of both worlds, the effect on battery life can be disastrous. Wireless communication
loses much of its attraction if your communicator requires a cable to connect to a power source.
Worse still while the power requirements of most electronic components are falling quickly, the
power required to radiate electrons through the air will remain substantial.
Secondly, as mentioned earlier, batteries are not improving at a rapid rate.
At this point, this may not seem to be a problem - simply limit the amount of data communi-
cated via wireless in order to save power. However, communication offers the ability to process
power hungry jobs remotely. Further, communication allows access to information and process-
ing unavailable on the local machine. Speech recognition, which often employs complex digital
signal processing algorithms, is not an efficient use of mobile processor resources. Access to a
wired remote server would improve (or enable) such applications.
As mobile platforms grow to include greater functionality (faster processors, larger disks, more
peripherals), their ability to perform resource intensive tasks grows as well. A resource intensive
task, in this context, includes any task which utilizes the processor extensively and/or calls upon
other system resources. In terms of power consumption, other resources (such as memory sub-
24
systems and spinning disks) are a major consideration. In fact, Denis Riley[12] showed that the
largest power reduction in the Smart Modules project was introduced by making the application
memory resident, thereby reducing disk accesses caused by swapping.
On the other hand, the thin client approach to computing assumes that all tasks are best per-
formed on a high performance server. The time required to send and receive the data, which
may be far greater than the time required to do the calculation, is ignored. Moreover, the power
required for such ransactions is also ignored. Unfortunately, this method leads to rapid battery
depletion.
For the purpose of this research, tasks performed using mobile devices are split into three
groups. The first group contains tasks for which wireless communication is not an option. Ex-
amples would be searching of the local filesystem or running an application while out of commu-
nication range. The second group includes tasks for which wireless communication is necessary.
Examples include browsing the Internet or sending email. The third group are tasks which could
be computed locally, or could be sent to a remote machine for processing. Examples of this in-
clude compression or decompression of files, or database queries where the complete database is
available locally and remotely. This third group is the concern of this report.
For this third class of task, how is it possible to determine quantitatively whether it is better to
process a task locally or remotely? Table 3.1 shows the tradeoffs between local and remote com-
puting. Both options present hidden power costs, as well as pros and cons which have nothing
to do with power consumption, but should influence any decision between local and remote pro-
cessing. The following sections develop an analytical model which determines whether remote
processing is appropriate for a task, based on several (power related) parameters.
25
Advantages DisadvantagesLocal Processing No transmission cost Less processing power
Lower latencyRemote Processing More processing power Security issues
Energy spent on transmissionProblems partitioning algorithm
Table 3.1: Local versus Remote Processing Advantages and Disadvantages
3.2.1 Power Consumption
The most obvious factors in the analytical model are the simple power consumption of the client
system and of the communication card. For purposes of the model, the power of the client includes
all of the subsystems except for the wireless interface. The energy required by a task computed
locally is expressed as
glocal : (Psystern-on q- Pcard-sleep) * Tsystern
where P~yst¢m-on is the power of the client when processing, Pcara-sl~¢p is the power of the com-
munication card in sleep mode, and T~y~t¢,~ is the time required to complete the job. Conversely,
the power required to communicate the task to a remote machine is expressed as
given that P~a~d-o~ is the power drawn by the communications card when transmitting, and Ttra~s
is the time spent transmitting data (both the input and results). Further, Ttr~ns can be expressed as
a function of the size of the transmission in bits, St~a~s, and the throughput in bits per second, ,-
Ttrans -=- ~Jtrans / T
26
It should be noted here that St~.,~s includes the packet headers, and that ~- should be measured
since the bandwidth stated by the manufacturer, and actual bandwidth utilization are usually
different. From a strict power consumption standpoint, if the energy required to communicate the
task is less than the energy required to compute it locally, then the task should be done remotely.
Otherwise it should be done locally. However, there are factors other than power consumption to
be considered.
3.2.2 Latency
Because the goals of many of these devices are consumer and real time applications, latency is
another concern in the decision of whether or not to process a task locally. In some cases the
server may be so busy with other processes that the latency to the end user becomes unacceptably
high. In other cases, the local machine may be more overloaded than the server, in which case
it makes sense, from a latency standpoint, to compute the task remotely (however, because of
transmission times, this is rarely the case).
For the purpose of the analytical model, these devices are considered to be single-tasking ma-
chines. As a result, while the task is being processed remotely the local machine is idling (and
therefore the wireless card is idling as well). This idle time can be accounted for in the model by
adding the energy used while the local machine is idling :
Note that Tsy~tem here is the time that the local machine spends processing the task. The server
is assumed to be of a similar speed as the client. This distinction will be clarified in Section 3.2.4
below.
27
3.2.3 Network Conditions
The very nature of the wireless medium may make communication overly costly. If the local
area is saturated with wireless packets, collisions become a significant problem. After retransmit-
ting several packets, a task which was a prime candidate for communication may consume more
power than had it been done locally.
Retransmissions can be added to the analytical model as follows :
~rernote -~ (Psystem-on q- Pcard-on) $ (T~rans q- ]~rpacket) q- ((Psystem-idle q- Pcard-idle) * ~[~systern)
where k is an integer number of retransmissions, and rpacket is the time required to re-send one
packet. Tpacket can be determined similarly to to Ttra~s as
Tpacket=~qpacket/T
where Spatter is the size of a single packet, up to 1472 bits. Note that this model represents a
best case scenario, that of a point-to-point link. Because of the way power management is im-
plemented for access points which connect to multiple clients, an unbounded wait is possible.
That is, when a client goes into sleep mode, it wakes up occasionally to see if the access point has
messages buffered for it. If more than one client has messages which have been buffered, only
one of them can receive its messages at a time. If the messages for the first client cannot all be
transmitted before the next beacon slot, the second client may have to contend for its messages
yet again. Thus, the unbounded wait. However, in a point-to-point system, each access point
only buffers messages for one client, therefore there is no contention for bandwidth. In a point-
to-point connection, the only things which will cause packet retransmission are hidden nodes and
28
radio-frequency noise.
3.2.4 Comparative Processing Power
Up to this point, the assumption has been made that the local and remote machines will complete
a task in the same amount of time, Tsystem. However, this is rarely the case. Often a wireless
client will be based on an embedded processor such as the StrongArm SA-1100 (e.g. the Itsy)
a Motorola Dragonball (e.g. the PalmPilot), while the server will have one or more high powered
processors, not to mention more main memory and faster secondary storage. As a result, the
server will generally be able to complete a task in a fraction of the time that the wireless client
would be able to. While this may not make a noticeable difference to the end user, it is fairly
significant in terms of power consumption. The ability of the server to process faster than the
client can be modeled as :
Tsystem )
where Tsystem is the CPU time had the task been computed locally, and the server is n times faster
than the client.
3.2.5 Threshold Calculation
The above equation is useful to measure the absolute power requirements of a task. However, in
the case of a system designer, it is more useful to rearrange the equation to calculate whether or
not a system under consideration could benefit from remote processing. Using a series of basic
algebraic methods, the equations for remote and local processing can be combined and simplified
29
to this equation
(Psystern-on if- Pcard-on) * (Ttrans q- kTpacket)
This equation will calculate the threshold value for computation time. That is, plugging into the
equation for a particular system and a task with a known data set size, the equation will produce
a time in seconds. If the task requires more than this amount of time to complete on the client, it
would benefit from remote computation. If a task can be completed locally in a shorter time, then
local processing would be more efficient.
3.3 Model Results
Using measured properties of the Itsy and the WaveLAN card, the model developed in the pre-
vious section was used to analyze when to process a task locally and when to send it to a re-
mote server. The Itsy draws approximately one watt when computing at full power (P~yst¢m-on),
and 450 mW when idle (P~yst¢m-idte). The WaveLAN card draws approximately 1.5 watts when
transmitting or receiving (P~ard-on), and a negligible amount (45mW) when sleeping (Pcard-sleep).
While the WaveLAN card specifications state a bandwidth of 2 Mbits/s, the throughput (q-)
closer to 1800 kbits/s. Figure 3.1 shows the results of the model derived in Section 3.2.5 when
applied to the Itsy. This figure depicts the thresholds at which the energy required for remote
computation is equal to the energy required for local computation. Above and to the left of each
threshold, the task would require less energy on a remote processor. Below and to the right of each
threshold, the task would be more energy efficient if computed on the local machine. The upper
chart shows the effects of a faster server machine. Intuitively, a faster server makes a greater dif-
ference on a longer task, and this is consistent with the information in the chart. The lower graph
30
5
4
3
2
1
500 1000 1500 2000 2500 3000 3500 4000 4500 5000
~12
~-4(
~ 2
O~500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Data set size (kilobits)
Figure 3.1: Local or Remote decision thresholds as calculated using the equation and measure-ments taken from the Itsy. The upper chart depicts the effect of increased server speed with respectto client speed assuming that there are no collisions. The lower chart shows the effect of packetretransmissions caused by collisions with packets from hidden nodes assuming equal client andserver speeds. Note that the baseline (equal speed, no collisions) line is the same in both upperand lower charts.
depicts the effect of one, two, and three packet retransmissions. Even one collision has a dra-
matic effect on the threshold, raising the required computation time by a factor of ~qp~c~et//~, or
810 milliseconds in the case of the Itsy.
Possibly the most significant pieces of information in this chart are the slopes of each line. Since
all of the thresholds are linear, the slopes represent a ratio of computation time to transmission size
at which the power consumption is the same locally and remotely. Thus, for example, the slope
of the threshold with equal processor speeds and no collisions is .0014 seconds per kilobit, or 1.4
milliseconds per kilobit. Power can be conserved by the remote processing of any task which
requires more processor time per kilobit transferred. Conversely, any process which requires less
31
No Collisions 1 Collision 2 Collisions 3 CollisionsEqual Speed 1.4 2.8 4.3 5.72x Speed 1.1 2.3 3.4 4.65x Speed 0.99 2.0 3.1 4.110x Speed 0.95 2.0 3.0 4.0
Table 3.2: Remote processing thresholds in milliseconds/kilobit
computation time per kilobit transferred is more efficiently dealt with locally. The slopes for each
of the lines in Figure 3.1 are listed in Table 3.2. This table also includes slopes not shown in the
graph, specifically when the processor speeds are unequal and there are collisions.
3.4 Results and Conclusions
Solving this equation using data from the Itsy shows that the most significant factors are transmis-
sion size and computation time. A likely candidate for remote processing will have, as is intuitive,
a small transmission size and a long computation time. Fortunately, there are several applications
which are common on handheld devices which fit this description. The first example would be
searching a contact list or calendar. While the transmission size is daunting if the entire file must
be transmitted each time, it is fair to assume that such lists are fairly static and can be stored both
locally and remotely (i.e. caching ). A second example is speech recognition. Sphinx II, a speech
recognition program being developed at Carnegie Mellon University, involves three distinct parts
[6]. The first is a fairly complex algorithm to create Cepstrum coefficients from raw audio data.
In so doing, the size of the data set is reduced by a factor of approximately 6 (from 31.25 kB/sec
down to 5.08 kB/sec). The second part is performs a vector quantization on the Cepstrum coef-
ficients, reducing the size of the data set yet again (from 5.08 kB/sec to 0.39 kB/sec) [6]. Thus,
an application of the equation presented above would be to determine whether to apply the Cep-
32
strum coefficient algorithm locally, or to transmit the raw audio data (either way, transmission
must take place since the final stage of the speech recognition is a database search which requires
more secondary storage than is generally available on a wireless device).
As with any analytical model, the results seen above are only as good as the measurements
used in the equation and the fidelity of the equation itself. This model accounts for several of the
most important factors in a point-to-point wireless link. Given that, if accurate data are collected
for a given system, the model derived above should generate accurate predictions as to whether or
not to compute a job remotely. Unfortunately, the difficulty comes in accurately quantizing factors
such as transmission size or computation time for a specific task. The transmission size consists
of the input set size and the output set size. The input set size is known, and the output size
can usually be estimated within an order of magnitude. Computation time can be approximated
based on knowledge of the task, user hints, or previous runs of similar tasks.
33
Section 4
Further Research
Currently, super capacitors are used in some high end, two-way pagers.[4] These devices have
a nominal load current on the order of 5 milliamps. However, a transmission from one of these
devices requires a current pulse of approximately 1A. A super capacitor is used to store energy
output from the power supply and pulse it to the communications hardware. These pulses have
a known maximum duration, and can be delayed if the capacitor has not fully charged since the
previous pulse (since the end user can not tell if a page has been delayed). The super capacitor
in this application allows for the use of a small power supply capable of only slightly more than
the nominal load. This arrangement would not work for CPU pulses, since they do not occur for
known durations, nor are they easily delayed if the capacitor is not at full capacity. If the CPU
demanded more power than the capacitor was capable of in this situation, the load would fall
to the power supply. Since this load would be roughly two orders of magnitude greater than
the rated capacity of the power supply, the supply would likely fail, possibly causing damage to
the system. However, a similar set up could possibly be used with the wireless communications
hardware on a mobile wireless device, since packets can be delayed in a way similar to pages, and
are of known duration. The down side of this method would be increased latency.
34
Today’s wireless devices are completely static in the way that they deal with communicated
data. Implementing the equation presented in Section 3.2 as a dynamic process allowing a device
to determine whether to process locally or communicate a task would be an interesting experi-
ment. Such an algorithm could be modified such that it could ’learn’ about its environment. For
example, a record of previous server performance could suggest how much faster the server is
than the client. A sliding window could be implemented which would use the running time of
previous tasks to approximate the running time for the process being evaluated.
35
Appendix A
Matlab code for Local/Remote Algorithm
function Tcomp = LocalRemote(S, k, n)
%This function returns a threshold computation time in seconds given a
%transmission size in kilobits, a number of packet retransmissions k, and a
%speed factor n. If the size parameter is a vector, Tcomp is
%a vector as well. The speed factor n is the number of times faster the
%remote server is when compared to the local machine. The machine dependent
%parameters are preset to measurements taken from the Itsy. Note that if
%the task requires greater than the threshold computation time, it is more
%power efficient to compute the job remotely.
Ptrans = 1.5 ; %Transmission Power (watts)
Pcomp = 1.0 ; %Local System On Power (watts)
Pidle = 0.5 ; %Local System Idle Power (watts)
B : 1800; %Bandwidth (Kb/s)
%Given size (Kb), determine comp time for ratio ==
ratioConst = (((Pcomp+Pidle) - (Pidle/n)) / ((Ptrans + Pcomp)/B));
for i = i : (length(S))
Tcomp(i) = (S(i)+(k*1472))/ratioConst;
end
36
Appendix B
Glossary
Available Capacity The total capacit~ Ah or Wh, that will be obtained from a cell or battery at
defined discharge rates and other specified discharge or operating conditions. [13]
C Rate The discharge or charge current, in amperes, expressed as a multiple of the rated capacity
in ampere-hours. [13]
Capacity The total number of ampere-hours or watt-hours that can be withdrawn from a fully
charged cell or battery under specified conditions of discharge.[13]
Cycle The discharge and subsequent or preceding charge of a secondary battery such that it is
restored to its original conditions.[13]
Duty Cycle The operating regime of a cell or battery including factors such as charge and dis-
charge rates, depth of discharge, cycle length, and length of time in the standby mode. [13]
Energy Density The ratio of the energy available from a cell or battery to its volume (Wh/L). Also
used on a weight basis (Wh/kg).[13]
Primary Battery A primary battery is a non-chargeable battery, such as an alkaline battery.
37