1. Jacob R. Lorch and Alan Jay Smith University of California, Berkeley
Reducing Processor Power Consumption by Improving Processor Time Management in a Single-User Operating System
Jacob R. Lorch and Alan Jay Smith
University of California, Berkeley

2. Welfare reform initiatives
Overall problem: not enough money for welfare recipients
Three possible initiatives
stop overpaying people who have filed fraudulent claims overstating their worthiness
stop overpaying people who have filed honest claims but who, through bureaucratic error, are being paid more than regulations require
reduce all payments by a certain percentage
Money : welfare recipient :: CPU time : process

3. Outline of presentation
Motivation
Background on MacOS
A different strategy for power management
Problems with this strategy and our solutions to these problems
Results of simulations
Conclusions

4. Motivation Power reduction is important on the desktop
expense, heat, fan noise, contribution to pollution
government regulation: Energy Star, Blue Angel
Power reduction is critical in portable computers
battery capacity per unit weight advancing slowly
limited space and weight available in portables
drives importance of power management
Importance of processor energy reduction
Accounts for large percentage of total power (18-34%)
Watching the CPU turning on and off, we got the impression it was frequently on when it wasn’t needed

5. Background on MacOS
Current power management (strategy “C”): inactivity timer
after 2 seconds of no activity and 15 seconds of no disk activity, processor turns off until next interrupt
Process scheduling: cooperative multitasking
processes switch out voluntarily
when a process switches out, it indicates a sleep period, the maximum amount of time it wants to remain unscheduled
if the process is busy, it will give a sleep period of zero

6. The basic strategy
Energy and time costs to turn on and off the CPU are low
Window-based operating systems are event-driven
We want the CPU off whenever it is idle
A process handles an event, then blocks
The basic strategy (strategy “B”):
Turn off the CPU when all processes are blocked, waiting for their next events
However, there are some problems with the basic strategy on MacOS...

7. The simple scheduling technique
Problem: MacOS sometimes schedules processes when they do not need to be scheduled
i.e. it wakes them up after too little time asleep
Why? Well, it is a single-user OS; who would care?
Technique “I”: never schedule a process whose sleep time has not yet expired

8. Sleep extension technique
Why use sleep periods: to perform periodic tasks
blinking the cursor, checking if time for backups
Not a real-time system, so deadlines not critical
in fact, current inactivity-based strategy freely violates these deadlines
Therefore, perhaps it is reasonable to let processes sleep for longer
Technique “S”: multiply all sleep periods by a multiplier s, set by user or OS to maximize energy savings while maintaining desired performance

9. The greediness technique
How processes should behave
receive event, process event, block for next event
perhaps, while waiting, do periodic tasks
How a process should not behave: “greediness”
failing to block, i.e. using sleep period 0, when not busy
happens often in MacOS
Technique “G”:
call a process greedy if it yields control g times without activity and without blocking
block greedy processes, but only for a certain period f, in case our heuristic is wrong

10. Trace-driven simulation
Trace collection (6 users): IdleTracer
only collects data while running on battery power
shuts off power management while tracing
Simulators: ItmSim (current strategy) and AsmSim (basic strategy)
Evaluation criterion 1: CPU energy savings
Evaluation criterion 2: Performance impact
estimated percent increase in workload completion time
increase comes from skipping over useful work when processor is off: when processor comes back on, such work must take place
heuristic: quantum is useful if it has activity or an event

11. Results 11.5% more battery lifetime than C
Performance impact for C was 1.84%, so that was the impact for BIS and BIG.
Performance impact for B and OPT was 0%.
Performance impact for BI was 1.08%.
11.5% more battery lifetime than C
20.0% more battery lifetime than C

13. Problems with sleep extension
Not as much energy savings per unit performance impact as greediness technique
does not even work well as a supplement to greediness
Does not eliminate an important problem with basic strategy: processes that fail to block
this is why even BIS did not beat C for user 2
Extending real-time sleep factors violates the user interface, and may produce undesirable effects

14. Future work Implementation of strategies
Collection of additional traces
Adaptation of techniques to other OS’s
Completely different energy-saving techniques for the CPU, to get closer to optimal (or even surpass it!)

15. Conclusions Basic strategy needs good CPU time management
Our suggested ways to improve CPU time management are:
never schedule processes that don’t want time
identify and block processes that are hogging the CPU
let processes sleep longer than they want to
With the best of these techniques, the basic strategy far surpasses the inactivity timer strategy
53% less processor energy consumption
20% more battery lifetime