Critical Power Slope: Understanding the Runtime Effects of
Frequency Scaling
Akihiko Miyoshi†,Charles Lefurgy‡,
Eric Van Hensbergen ‡, Ram Rajamony ‡,
Raj Rajkumar †
† Real-Time and Multimedia Systems LabDept. of Electrical and Computer Engineering
Carnegie Mellon University
‡Austin Research LaboratoryIBM
The Question
• Operating Points – [600MHz,6V], [525MHz,4.2V],[450MHz,2.8V],[375MHz,2V],
[300MHz, 1.7V], [225MHz,1.5V],[150MHz,1.45V]
• Where should I operate (for energy efficiency)?– Dynamic Voltage Scaling (DVS) algorithms– Lowest performance without sacrificing user/application
requirement
• Why lowest performance is not always the best– Even for voltage scaling systems
Energy Efficiency
...
power
time
activeE
t
Watts
activeE
idleE
t
Watts
Low frequency High frequency
• Majority of OS policies assume
• Not always the case!– When it is not the case?
– How do we determine this?
Assumption
<activeE
t
Watts
activeE
idleE
t
Watts
• Motivation– < : not always true– How do we choose which operating points to use?
• Measurement results• Analytical model: Critical Power Slope• Analysis on voltage scaling systems• Conclusion
Outline
lowfE highfE
Power Management Techniques
• Provides multiple operating points– [600MHz,6V],[450MHz,2.8V],[300MHz, 1.7V]…etc
• Three empirical data points– Frequency Scaling
• PowerPC 405GP
– Clock Throttling• Pentium with ACPI
– Voltage Scaling• Strong ARM SA-1100
• Note: We are not making any statement on the benefits of these techniques! – These are merely samples which real systems use to manage power.
Basic Results• Runtime and frequency
– CPU intensive workload: inverse relationship
• Power and frequency– Frequency scaling, clock throttling processors
• CPU active: linear relationship• CPU idle: constant
m: slope
CPU active
CPU idle
Power
Frequency
Energy Consumption
• Compare energy consumption at different operating points– Same workload W – Same amount of time t
activeE
idleE
tpower
time
0
500
1000
1500
2000
2500
3000
12 25 37 50 62 75 87 100
CPU performance (%)
J oules
Extra IdleSystem Active
Energy consumption (Pentium L1 cache read hit)
2490J
2591J174.3sec
Energy consumption (PPC L1 cache read hit)
0
50
100
150
200
66 133 200 266MHz
J oules
Extra IdleOthersSDRAMCPU
136J66.4sec
162J
Measurement Results
• Results consistent with different workloads– Register, L1 cache, memory, disk accesses– Web server (Pentium)
• Pentium– Highest frequency always energy efficient
• PowerPC– Lowest frequency always energy efficient
• Why?– What happens on voltage scaling systems?
• Motivation– Which operating points should we consider?
• Measurement results– Pentium: highest performance better– PowerPC: lowest performance better
• Analytical model: Critical Power Slope• Analysis on voltage scaling systems• Conclusion
Outline
• CPU intensive workload W• Frequency
– Assume utilization of system = 1– units of time to complete W– Energy consumed
• At frequency– Time to compute W:– Remaining extra idle time:
Characterization
minf
minfT
minminmin fff PTE )( minff
ff
fT minmin
)min
1min( ff
fT
idleff
ffff
ff PTPTE )1()( minmin
minmin
– Power increases linearly with frequency– m: slope
• Is energy efficient??– True if – Depends on m
Critical Power Slope
)( minmin ffmPP ff
idleff
ffff
ff PTffmPTE )1()]()[( minminminmin
minmin
minfminff EE
• Use slope m to characterize system– Find hypothetical m for and call it
Critical Power Slope (CPS)
Critical Power Slope cont’d
minff EE
min
min
fPP
criticalidlefm
What does it mean?
Freq
Power
criticalm
minf
idleP
min
min
fPP
criticalidlefm
minfPidleP
criticalmm
criticalmm
• If– Energy efficient to run at higher freq.– Pentium
• If– Energy efficient to run at lower freq.– PowerPC
Implications of CPScriticalmm
criticalmm
028.%5.128481215
MHzWWcriticalm020.%5.12848
1530 MHzWWm
0038.6602.227.2 MHzWWcriticalm0043.66266
27.213.3 MHzMHz
WWm
minff EE
minff EE <
J.Pouwelse, K.Langendoen, and H. Sips, “Dynamic Voltage Scaling on a Low-Power Microprocessor”, MOBICOM2001
Voltage Scaling Processors (Strong Arm SA-1100)
• Look at every operating point at frequency
• If– Energy efficient at higher frequency than
• If– Energy efficient at lower frequency than
CPS for voltage scaling system
xf
fxPP fx
idlefxfx
criticalm
fx
critical
fx mm
fx
critical
fx mm xf
xf
Analysis on SA-1100
• Above 74MHz
• At 74MHz
• Below 74MHz
• Energy Inefficient below 74MHz!
001.0744612174
MHzmWmWMHz
criticalm
001.0597410612174
MHzMHz
mWmWMHzm
fx
critical
fx mm
fx
critical
fx mm
SummaryPower
Frequency
Power
Frequency
Power
Frequency
Pentium PowerPC
SA-1100 CPS: Characterizes the runtime trade-off of power management techniques
Conclusion
• Which operating points should we consider?– Traditional DVS algorithms attempt to go to lowest
frequency– Not always the best choice
• Critical Power Slope • Identifies energy inefficient operating points• Can be used to inform OS (DVS algorithms) of operating
points it should not consider