1. Hui Chen, Shinan Wang and Weisong Shi Wayne State University
Where Does Power Go in a Computer System: Experimental Analysis and Implications
Hui Chen, Shinan Wang and Weisong Shi
Wayne State University

2. Outline Introduction Power Measurement Evaluation Implications
Conclusions

3. Introduction The power problem of computer systems
Restrict performance improvement
Frequency limitation
Hard to cool down
Influence user experience
Battery lifetime of mobile devices
Waste a large amount of energy consumption
Energy un-proportional in data centers
Low utilization VS high energy consumption
As we know that, power consumption recently acquires more and more attention.
Why this is a problem?

4. Introduction (cont.) Methods to solve this problem
Low-power circuits design
New material (PCM)
Power-aware system design
Hardware supplies different power states
Operating system makes power-aware strategies
Power-aware software & application design
Optimize during code compiling
Decrease application performance requirement when battery level is low

5. Motivation
Computer system still consumes an ever increasing amount of energy
The idle power does not decrease too much
Idle power is not used for computing
Account for a large amount of the total power dissipation
One of the main reason that cause energy un-proportional
Where does power go in a computer system?
Even though a lot works have been done in the last several years…
Idle power is the power dissipation when the system is in idle state.
Usually, we think this part of power is not used for computing. Thus, many people try to decrease the idle power.

6. Outline Introduction Power Measurement Evaluation Implications
Conclusions
In able to answer this question, we first need to measure the power.

7. Experiment Platform Two desktops of different periods
We use two desktop computer to do the experiment.
The first one was produced in 2005 and the second one was produced in 2010.
2.6GHz
Two desktops of different periods
The same producer
Single core VS multi-core
DDR VS DDR3
Different disk size

8. Power Measurement Direct Method Indirect Method (subtract) CPU, Disk
Memory, NIC
We connected a resistor for each power cable and measure the voltage.

9. Indirect Power Measurement
Execute a benchmark application to stress a component
Find out with cables supply power for each device
Which cables’ voltage change.
Example: the three brown cables are related with memory.

10. Indirect Power Measurement (cont.)
#define CACHELINE_SIZE 64
#define L2CACHE_SIZE 2048
#define ARRAY_SIZE (L2CACHE_SIZE * 1024/CACHELINE_SIZE * 2)
typedef struct{
int data[CACHELINE_SIZE/4];
}LINE;
LINE array[ARRAY_SIZE]; …
unsigned int size = ARRAY_SIZE;
unsigned int i = 0;
while(1){
array[i%size].data[0] = i;
i++; }
By setting up the size of the LINE data structure and the ARRAY_SIZE, we guarantee the data is always read from memory in each iteration.

11. Outline Introduction Power Measurement Evaluation Implications
Conclusions

12. Idle Power 3M
Others
Total
PC05
19.2W
26.1W
45.3W
PC10
12.7W
28.3W
41W
While we trying to decrease the idle power of part components, we increase the idle power of other components.
The total idle power does not drop too much.
The other part is mainly consumed by chips on mother board.

13. Idle power (cont.) The idle power of CPU and memory dropped a lot.
Similar situation was not shown on disk.

14. Idle Power (cont.)
Even though the power of processor and memory have decreased, the power of chips on motherboard increased.
PC05
PC10
Other components should acquire similar or even more concentration .

15. The active power of CPU The active power decreased about 5 – 22Watts.
We execute 5 benchmarks to stress CPU.
100% CPU utilization.
The active power decreased about 5 – 22Watts.
The active power is significantly different when executing each benchmark, even though all the utilization is 100%.

16. Outline Introduction Power Measurement Evaluation Implications
Conclusions

17. Not suitable to be used for power modeling.
CPU Utilization
CPU Utilization is not a good indicator of power dissipation.
When CPU utilization is 100%, the difference of power may be more than 10W.
Not suitable to be used for power modeling.
We need to find other indicators that could reflect power dissipations.

18. Controllable Cache Size
Cache benchmark generates much more power than memory benchmark.
Through control cache size we could control the power dissipation of CPU.
If we decrease cache size, this forces CPU to read data from memory.

19. An example
High
Indicator (battery level)
Low
Cache

20. Power Transform Efficiency
DC Power
AC Power
Transform Efficiency =
The transform efficiency of AC to DC is low.
High transform efficiency is required in able to save power.

21. Multi-core Task Allocation
Caused by global frequency and resource share.
The power of CPU does not decrease too much when idling part of cores.
From the energy efficiency point of view, idling part of cores while make the other parts busy does not save power.

22. Conclusions We measured the power of several main components.
The total idle power does not decrease too much.
The idle power of other parts and disk should acquire more attention.
From the experiment result, we derive several implications that are important for power-aware system design.
CPU utilization is not a good indicator of power dissipation.
Cache size should be controllable.
Power transform efficiency should be increased.
Idling part of cores does not save power.

23. More question?
Thank you!