Zhou Peng, Zuo Decheng, Zhou Haiying Harbin Institute of Technology 1.

Slow Down or Race to Halt: Workload effect on

Energy Effective

Zhou Peng, Zuo Decheng, Zhou HaiyingHarbin Institute of Technology

1

Harbin institution of technology

1.Introducation 2.Workload effect on Energy effective 3.Conclusion & Future works

Outline

2


Background

Green computing is imperative

Increasing of computers

Increasing of energy cost

Increasing of Carbon emissions

3


Energy effective

Moore’s law Moore’s law for energy effective

4


Explosive growth of the tasks and complexity

Linear growth of energy density in battery

Exponential growth of code ； e.g. Linux code in tar.gz format increase from 117K(0.11) to 109M(3.11.1)

Explosive growth of applications; e.g. apps for android and apple

Explosive growth of amount of computation; e.g.AI & Big data

Linear improve of battery

5

VS

Battery life become shorter and shorter ； e.g. smart phones


Main technologies to improve energy effective◦ Hardware level: Low power devices◦ System level: Power-management mechanisms in different

levels◦ Application level: Consolidate with virtualization

Power-management mechanisms◦ Circuit level: Clock-gating◦ System level: DPM◦ Processor level: DVFS/DFS/DVS, C-state

Motivation

6

To Shutdown unused component or circuit


According to the present researches:◦ C-state can save up to 44%[1] energy◦ DVFS can save 13%[2] to 70%[3] energy

Limitation of present research◦ All the results come from particular system with special

application or SPAC CPU. ◦ Few works can consider the effect of workload to the energy

consumption.

7

Slow Down or Race to HaltDVFS vs. C-state


Two solutions: slow down & race-to-halt

Objectives: To evaluate the energy effective of DVFS & C-state with different task models

8

DVFS vs. C-state

Slow down race-to-halt

Typical technology DVFS C-state

Runtime power Dynamic & low Higher

Time to finish task Longer short

Deadline miss High risk Lower risk

Energy effective Save lots of energy Save lots of energy

DVFS vs C-state: which is better in energy effective?



Outline

9

Harbin institution of technology 10

Relate works & Premise

2dd staticP CV f P

2( ) /dd t ddf k V V V

Relationship of the power and the frequency:

Relationship of the voltage and frequency:

k: is a circuit dependent constant Vt: is the threshold voltage

C : is the capacitance of the transistor gates f : is the frequency Vdd: is the supply voltage of the device. Pstatic: represents power consumed from leakage

mechanisms.

, Note that: The operation frequency almost has a linear relationship with voltage.BUT, decreasing the frequency and keeping the voltage constant does not contribute much to energy saving. It just saves the cost of cache misses[11] .


DVFS Modeling ◦ Defining the amount of computation/ instructions for a

task/workload is W, ◦ and then within a period of run-to-completion, the energy

consumption of task is

is energy consumption based on dynamic power

is energy consumption based on leakage power

Summary:◦ DVFS: compute the energy consumption of processor but

ignore the energy cost of cache misses.

DVFS model

11

22

( )( )

dd staticd dd dd

dd t

WV PE V PT CV W

k V V

C: capacitance f : frequency Vdd: runtime voltage Pstatic: leakage power Vpeak: peak voltage Tr: Time to finish task Ts:Time to sleep W: workload, the instruction

cycles of a task

Tr+Ts = W/fd

2ddCV W

2( )dd static

dd t

WV P

k V V


C-state Modeling ◦ Defining the amount of computation/ instructions for a

task/workload is W, and then within a period of run-to-completion, the energy consumption of task is

◦ Tr+Ts is the interval time of a task run-to-completion based on DVFS

Tr+Ts = W/fd Summary:

◦ C-state operates at higher voltage, So C-state finish a task faster than DVFS.

◦ If all the tasks is completed, system changes to sleep mode.◦ is very low, which can be ignored.

C-state model

12

2c peak static r sleep sE CWV P T P T

C:capacitance f :frequency Vdd: runtime voltage Pstatic: leakage power Vpeak: peak voltage Tr: Time to finish task Ts:Time to sleep W: workload, the instruction

cycles of a task

sleepP


The derivative of energy model

13

Analysis of the optimal voltage

2 3

2( ) 2

( ) ( )st dd st

dddd t dd t

WP WV Pdv E CV W

k V V k V V

The extreme point in energy model shows that◦ Workload W is not the key influence factor to the minimal

energy consumption ◦ The minimal energy consumption is only depended on the

characteristics of devices

In order to minimize the energy consumption and also try to find the best voltage, we can get the derivative of energy models


C-state becomes popular because Pstatic (leakage power) increase effects

We can consider time t as the workload arrival time, when , rewrite the equation

14

Workload effect on Energy effective

2 2( ) ( ) ( )d c dvfs peak staticdvfs

WE t E E CW v V P t

f

In order to evaluate the energy effective of DVFS and C-state ，We get the difference value of the two energy models ：

( ) 0E t

2 2( )peak dvfsst dvfs

CW Wt V v

P f


For Poisson distribution workload◦ The average arrival rate of task is λ0;

◦ The average interval time of task is t=1/ λ0

Summary:◦ DVFS and C-state save the same energy in this situation

When deadline tdeadline < t, C-state saves more energy than DVFS;

◦ When the arrival rate λ>λ0, DVFS is better than C-state


15

2 2

0

1( )peak dvfs

st dvfs

CW WV v

P f

Harbin institution of technology 16


For Periodic distribution workload◦ C-state saves more energy if and only if the deadline is

smaller than period, i.e. tdeadline < t;

◦ DVFS does not shutdown the processor after the task finished.



Outline

17


Evaluate the energy effective of DVFS & C-state with different task models◦ The most energy saving voltage is only depended on the

characteristics of the device itself. ◦ The energy effective of DVFS and C-state is closely related to

the arrival rate of the tasks and the features of workloads. ◦ For the heavy workload systems, DVFS is better in energy

saving than another. The result is consistent with the conclusion in [5].

Conclusion

18


In this paper, we mainly focus on processor and ignore the energy consumption during state transition.

So, future works will be:◦ To analyze the effects of cache hit rate on energy effective in

the whole system.◦ To take the reliability into consideration. ◦ To explore the schedulability analysis methods for the energy

and reliability critical system.

Future works

19


1. Pavel Somavat. Accounting for the Energy Consumption of Personal Computing Including Portable Devices

2. Rotem, E., et al. Energy Aware Race to Halt: A Down to EARtH Approach for Platform Energy Management. Computer Architecture Letters.

3. Shekar, V. and B. Izadi. Energy aware scheduling for DAG structured applications on heterogeneous and DVS enabled processors.

4. Valentini, Giorgio Luigi, et al. An overview of energy efficiency techniques in cluster computing systems.

5. Petters, S. M. and M. A. Awan., Slow down or race to halt: Towards managing complexity of real-time energy management decisions.

6. Awan, M. A. and S. M. Petters. Enhanced race-to-halt: A leakage-aware energy management approach for dynamic priority systems. Real-Time Systems

7. Naik, R. Biswas, S. , Datta, S.; Distributed Sleep-Scheduling Protocols for Energy Conservation in Wireless Networks. System Sciences,

8. Le Sueur, Etienne, Heiser, Gernot. Dynamic voltage and frequency scaling: The laws of diminishing returns.

9. Le Sueur, E. and G. Heiser. Slow Down or Sleep, that is the Question.

10. Schmitz, M.T., et al.; Energy-Efficient Mapping and Scheduling for DVS Enabled Distributed Embedded Systems.

11. Wan Yeon Lee. Energy-Saving DVFS Scheduling of Multiple Periodic Real-Time Tasks on Multi-core Processors.

12. F. Paterna, et al.Variability-Tolerant Workload Allocation for mpsoc Energy Minimization under Real-Time Constraints

Reference

20


Thank you!

Zhou Peng, Zuo Decheng, Zhou Haiying Harbin Institute of Technology 1.

Documents

energy consumption of

minimal energy consumption

energy dvfs vs cstate

task slide

derivative of energy

time t

circuit slide

energy cost of cache