1. Jacob R. Lorch Microsoft Research
Operating System Modifications for Task-Based Speed and Voltage Scheduling
Jacob R. Lorch
Microsoft Research
Alan Jay Smith
University of California, Berkeley
First International Conference on
Mobile Systems, Applications, and Services
May 7, 2003

2. Dynamic voltage scaling
The ability to quickly and efficiently change CPU supply voltage without rebooting
Appearing on commercial CPU’s (Transmeta, AMD, Intel)
Reducing voltage reduces energy consumption – E α V2
Lowering voltage necessitates lowering speed
When is lower energy worth the lower speed?
Operating System Modifications for Task-Based Speed and Voltage Scheduling

3. Interval-Based Scheduling
First proposed by Weiser et al [WWDS ’94]
Used commonly today (LongRun, PowerNow!, Enhanced SpeedStep, Windows XP)
Approach
Divide time into intervals
Set speed for upcoming interval based on recent CPU utilization
Problem: recent usage says nothing about actual performance requirements (i.e., deadlines)
Especially true for one-shot tasks, e.g., responding to user interface events
Operating System Modifications for Task-Based Speed and Voltage Scheduling

4. Task-Based Scheduling
Task-based schedulers take into account what tasks are running
Knows or estimates their deadlines and performance requirements
From this deduce minimum required speed
Task-based schedulers have been built
Pillai et al [PS ’01] used knowledge of task characteristics in a hard real-time environment
Flautner et al [FURM ’00, FRM ’01] estimates when tasks begin and end, determines what speed is necessary to meet estimated deadlines
Operating System Modifications for Task-Based Speed and Voltage Scheduling

5. RightSpeed: Our Task-Based Scheduler for Windows 2000
Applications
instrumented
to describe
their tasks
Uninstrumented applications
Task information inference via
user-interface event interposition
System calls to describe task boundaries and deadlines
Operating system routines for scheduling speed and voltage
CPU capable of dynamic voltage scaling
Implemented without OS source code access
Novel automatic task detection system
Implements PACE scheduling algorithm
Operating System Modifications for Task-Based Speed and Voltage Scheduling

6. Outline Automatic task detector
PACE approach to dynamic voltage scaling
What it is
How processor and system realities necessitate changes to PACE implementation
Implementing RightSpeed on Windows 2000
Experimental results
Recommendations for processor design to enable greater energy savings
Operating System Modifications for Task-Based Speed and Voltage Scheduling

7. Automatic task detector
Inferring task start
Use Windows message hooks to see user-interface events
Preserve task type information to better predict future tasks
Inferring task completion
Old approach: track set of threads responsible for event
Infer task end when all threads are blocked and no I/O
Found to produce the same results most of the time
Implementation uses low-priority thread and I/O counter
Task: everything system must do in response to a user-interface event
Task information database
Task
type 1
Task
type 2
Task
type 3 …
Sample of
recent task CPU
requirements;
speed schedule
How fast should I run?
I took this long.
Task starts
Task ends
Operating System Modifications for Task-Based Speed and Voltage Scheduling

8. PACE Task-based scheduling traditional approach Speed (MHz)
Determine speed needed to meet deadline… saves energy by running no faster
If task likely short, decrease performance… saves more energy by running slower
PACE goes one step further
Start a task running slowly, and speed up as the task progresses
Performance is the same, but since tasks are often short you save even more energy by usually running even slower
From distribution of recent similar tasks, can calculate best way to vary speed: s3 Fc(w) = K
1500
Speed (MHz)
1000
500 25 50
Time (ms)
Constant
schedule
PACE
schedule
Operating System Modifications for Task-Based Speed and Voltage Scheduling

9. Computing PACE schedules
Obtain sample of recent tasks’ work requirements
Weight more recent ones more heavily
Infer distribution from sample
Compute schedule from distribution and CPU properties
Task Kc
Task Kc
Probability
Speed
Task Kc
Task Kc …
Time
Cycles
Distribution
estimation
s3 Fc(w) = K
Operating System Modifications for Task-Based Speed and Voltage Scheduling

10. Challenges of implementing PACE
Energy not proportional to speed squared
More general equation s2 E’(s) Fc(w) = K
Power curve not concave up
Ignore certain settings
Limited set of settings
Round to nearest speed
Limited timer granularity
Reduce from 10 ms to 1 ms
Round schedule phases
I/O wait time
Accelerate after I/O
Requires I/O start/stop notification
Operating System Modifications for Task-Based Speed and Voltage Scheduling

11. Reducing computation time for PACE
Pre-computation
Limited set of speeds so small set of valid schedules
Similar values of K yield the same schedule
Formula only used when RightSpeed is installed
Binary search for least-energy schedule meeting deadline
Only compute when idle
Uses low-priority thread K
(“performance level”)
Operating System Modifications for Task-Based Speed and Voltage Scheduling

12. RightSpeed on Windows 2000
Task manager driver augments OS with interface allowing applications to describe task information
Virtual file system interface
RightSpeed library exports task-based interface
Infer task information from oblivious applications
Automatically load RightSpeed library into each process
Interpose user-interface event delivery with message hooks
Generate task calls on behalf of unmodified applications
Infer completion of tasks
use low-priority thread to know when all threads are blocked
use filter drivers to determine when I/O is ongoing
Compute PACE schedules in low-priority thread
Operating System Modifications for Task-Based Speed and Voltage Scheduling

13. Architecture of RightSpeed
Application 1
Application 2
Application 3 …
RSLib
RSLib
RSLib
User Mode
Kernel Mode
RSInit
RSIoCnt
RSLog
RSTask
Virtual File System Interface
Task Type Group (TTG) File Manager
Task Manager
Sample Queue
Idleness Detection
Timer Resolution
Automatic Schedules
Speed Controller
Operating System Modifications for Task-Based Speed and Voltage Scheduling

14. Experiments
Operating System Modifications for Task-Based Speed and Voltage Scheduling

15. Performance of RightSpeed
Filtering I/O operations adds 0.3 — 1.5% overhead
OS support for this simple operation would reduce this
Decreasing timer granularity adds 0.7 — 1.5% overhead
Overhead to perform RightSpeed operations on the order of a few microseconds
Schedules achieve goals with high accuracy even though OS is not real-time
Delay past deadlines within 0.5% of target
Number of deadlines missed within 1.5% of target
PACE schedule computation efficient
4.4 μs on AMD, 17.2 μs on Transmeta
But… PACE energy savings virtually non-existent!
Operating System Modifications for Task-Based Speed and Voltage Scheduling

16. Efficiency of CPU settings
Traditional DVS model
Power a cubic function of speed
Concavity of power vs. speed
Makes intermediate settings very worthwhile to use
Efficiency of a setting: savings from using that setting
Example
1 GHz setting saves ~50% vs. using 500 MHz and 1.5 GHz
If task is short, 500 MHz / 1.5 GHz schedule best; otherwise 1 GHz is
Without concavity, intermediate speeds unhelpful
1500
1 GHz
Speed (MHz)
1000
500 MHz /
1.5 GHz
500 25 50
Time (ms)
Operating System Modifications for Task-Based Speed and Voltage Scheduling

17. Processor CharacteristicsTM
2 of 5 speeds w/ efficiency < 0.5%
Mobile Athlon 4
3 of 5 speeds w/ efficiency < 1.5%
2 of those negative, so never useful!
Centrino
3 intermediate speeds have 3.7%, 6.0%, 9.8%
1 speed with negative efficiency
To benefit from intermediate speeds, processor design should value not just the slope, but also the concavity.
Operating System Modifications for Task-Based Speed and Voltage Scheduling

18. High Speed… in Mobile Processors?
Current trend in mobile processors is to reduce the minimum and maximum speed
Prolonged use of high speeds can damage chips
Improper use of high speeds can tank battery life
How can PACE help?
Having more speeds available can’t increase energy consumption because PACE computes optimal schedule
With PACE, energy consumption may decrease — it can use availability of high speeds to reduce the speed tasks start at
Only in rare cases when tasks happen to be long are high-power settings necessary, and then only for milliseconds
Operating System Modifications for Task-Based Speed and Voltage Scheduling

19. Increasing maximum CPU speed
Min speed 200 MHz, max…
Simulation varying maximum speed of CPU
Past/Peg gets worse, but Stepped and PACE improve
PACE energy consumption goes down 19.5% by having higher speed available
Conclusion: Having high-power settings can save energy with proper scheduling
Operating System Modifications for Task-Based Speed and Voltage Scheduling

20. Conclusions
Task-based scheduler can be built on Windows 2000 despite lack of source code
Automatic task detector does not require complicated OS interposition
Task-based scheduling and PACE calculations can be done efficiently
Saving energy using intermediate speeds requires concavity in power vs. speed curve
Adding a higher speed to a processor can reduce energy consumption, given intelligent scheduling
Operating System Modifications for Task-Based Speed and Voltage Scheduling