Transcript Slide 1
Power Management in SDR
Max Robert, Jeffrey H. Reed
Mobile and Portable Radio Research Group (MPRG)
Virginia Tech
September 14, 2004
1
Overview
Power fundamentals
Overview of approaches
Current state of technology
Power management for SDR
Operation states
Interface descriptions
Conclusion
2
Power Basics
P CV f
2
Terms:
Fixed attributes
α: switching activity
C: capacitance
V: voltage
f: operating frequency
Switching activity (algorithm-specific)
C is fixed
Attributes open to modification
V, f
3
Software-controlled power
Some attributes are determined at design
time and cannot be changed at run-time
Compiler optimizations
Waveform design
Attributes that can change at runtime
Operating voltage
Operating frequency
Timing control
Thread management in the case of processors
Active components
4
General Power Management
Power management split into three principal
categories
-
Previous work on each section varies in depth
T h re a d p rio rity a lg o rith m s
D yn a m ic Vo lta g e S ca lin g (D V S ) a lg o rith m s
D yn a m ic F re q u e n cy S ca lin g (D F S ) a lg o rith m s
P o licy se le ctio n a lg o rith m s
A p p lica tio n
- A d va n ce d C o n fig u ra tio n a n d P o w e r In te rfa ce (A C P I)
- O p e ra tin g S yste m P o w e r M a n a g e m e n t (O S P M )
O p e ra tin g S yste m /E n viro n m e n t
H a rd w a re
-
M u lti-vo lta g e H W (C ru so e )
M u lti-fre q u e n cy H W (C P U s, F P G A s, A S IC s)
F le xib le R F
E fficie n t co m p ile rs
A lte rn a te d a ta flo w s
5
Software-Controlled Attributes
Timing management
Thread priority in the case of a GPP or (sometimes) DSP
Bus/message management in system
Algorithm may optimize wait times to cluster work for component
Voltage and Frequency selection are related
Higher voltage will allow higher frequencies
Optimal voltage for frequency not necessarily the best choice
Active component selection is a subset
Voltage switching may be slower than frequency switching
May desire to maintain operating range for quick response
Set voltage or frequency to zero for that component
Flexible RF
Still unclear what attributes of the RF will be software-controlled
Mixer bias, filter BW, others
Framework- and application-based strategies need to be sufficiently
flexible to allow smooth integration of flexible RF control
6
Application & Hardware
Significant previous research
Adaptive management algorithms
Advanced power hardware-level power
management techniques
Application- and HW-based strategies well
suited for static applications
Current way of developing power-saving
strategies
Fixed waveform
Can be optimized to specific platform
7
Operating System/Environment
Software structure necessary to support
power management functionality
Standard interface
Switch between different states
i.e.: sleep (several levels), active
States not necessarily limited to sleep modes
Standard management structure
Maintain state of all devices in system
State machine for describing system
Unified structure for handling associated
devices
8
State-of-the-Art
Development limited to PC needs
BIOS-based management
Power management for laptops
Sleep mode management
BPM (BIOS power management)
Has no awareness of the user’s (or
application’s) needs
Operating-system based management
OSPM (OS power management)
Current de-facto standard
Most publications today are algorithms for the
efficient switching between states using OSPM
9
ACPI
Advanced Configuration and Power
Interface
State machine used to describe
machine configuration
States associated with different parts of
the system
Common interface provided to enact
changes in the state of the system
10
ACPI States
Basic set of
states
Cx
CPU states
Dx
States for
peripheral
device
Modem
Network card
Screen
Hard drive
11
Interface Descriptions
Multiple standardized interfaces
provided
Example
AcpiEnterSleepStatePrep
AcpiEnterSleepState
AcpiLeaveSleepState
Provides common interface for the
change of states for the system
12
OSPM
Operating System Power
Management
Model describing partitioning of power
consumption management
Operating system determines when to
trigger power management features
BIOS determines how to perform power
management features
13
OSPM/ACPI
OSPM and ACPI
integrated
OSPM provides
mechanism for
selection of
mode
Kernel
initiates
action
ACPI provides
common
interface to
hardware
14
Power Management For SDR
SDR places challenges different from
classic communications system
Can support application swapping
Needs to support wide set of devices
Variety of needs and states
Difficult to narrow to small, well-defined set of
states
Requires sophisticated power control
structures
Applications can be more predicable than PC
Possible to determine “fast enough” speed
Blind throttle for the application may not be
enough
15
State Support
ACPI supports mesh state machine
Assumes basic device states can be throttled
S1
S2
Sn
Linear transitions (throttle) are a subset of the
mesh state machine
S1
S2
Sn
16
Problems with Mesh SM
Assumes that all transitions are fundamentally
“equal”
Does not take into account QoS issues related with
state change
Example:
Voltage and frequency are fundamentally linked
Increased voltage will allow a higher set of
frequency settings to be supported
Throttle transitions based on the assumption that lowest
possible voltage is supported for the desired frequency
If a change in voltage incurs a higher time delay
than a change in frequency, could lead to unplanned
additional latencies
17
Rate-Change Support in
Communications
Example (802.11b):
Support alternate processing speeds for different
sections of received frame
P L C P P re fix
P L C P P a y lo a d
11 M b p s P S D U
1 M b p s (P re a m b le + H e a d e r)
CW
192us
Tra n s itio n
Tra n s itio n
P ro ce s sin g
ty p ic a lly ~ 4 0 0 u s
Fast
S lo w
Benefits
D e c is io n p o in t: d is c a rd fra m e ?
Minimizes required computing power
Provides ability to discard frame before high-speed
processing is necessary
18
Rate Change and SDR
Waveform takes place of “user” in SDR
Latencies associated with change of state need to be
taken into account
State switching needs to be in order of microseconds
Millisecond-level switches may be too slow for some
waveforms
Ideally, should cluster state changes into transition state
Example:
Crusoe TM5400 automatically controls voltage and
frequency settings
Slow ramp in voltage for up-frequency changes followed by fast
frequency change
Fast down frequency change followed by slow voltage change
Changes performed automatically
Possible for some equipment to leave change requests up to
the application
Voltage regulator can have a significant impact on the
transition speeds in core operating voltage
May be too slow (ms+) for some waveforms
19
State Machine Description
Break down state
machine into slowchange states and
related fast-change
states
Provides application
with ability to change
states quickly during
waveform operation
Also supports sleep
or standby
operation
F 1 ,1
F 1 ,2
F 1 ,3
V1
F 2 ,1
F 2 ,2
F 2 ,3
V2
F 3 ,1
F 3 ,2
F 3 ,3
V3
20
Sample Operation
Fast operation
Can cycle between 500 and
700 MHz
700
1 .8 V
May choose not to
transition, since change to
600 or 700 MHz expected
soon
300
400
500
1 .5 V
100
200
300
1 .2 V
Can still transition to lower
powers
600
500 MHz may be more
efficient at 1.5V
500
Support significantly lower
power consumption levels
Same concept can apply to
other devices
FPGAs, ASICs, CCMs, DSPs
21
Common Interface
Design of common interface will have to
wait until conceptual framework is
finalized
Will rely on ACPI to determine appropriate
interfaces
Will also rely heavily on SCA 3.0 interface
specifications
SCA 3.0 concentrates on non-CORBA interface
descriptions
Challenging task
Generic nature of hardware makes static
definition of interfaces unlikely
Will most likely require a generic structure
May be able to leverage AML
22
Application-Level Power Management
Algorithm development
Field of research currently has large number of
contributions
Primarily concentrating on PC-based systems
ACPI/OSPM
Clear from OEPM that SDR will have some
unique characteristics
Optimization strategies will be based on the
permutations possible by conceptual
framework
This research venue cannot proceed until
conceptual framework is complete
23
Conclusion
Some concepts in power management are
fairly mature
Current state-of-the-art does not cover all
needs of SDR
PC power management
Voltage and frequency scaling
Policies and algorithms
Unique issues related to nature of SDR
Actively developing techniques to resolve
these issues
24
Acknowledgement
This work is funded by the DCI
Postdoctoral Research Fellowship
and the MPRG Affiliates Program
25