Energy Model for Multiprocess Applications Texas Tech University

advertisement

Texas Tech University

Energy Model for

Multiprocess

Applications

Texas Tech University

Faster Computers = More Energy

• Moore’s law predicted 2 fold yearly increase in transistor count for inexpensive devices

Transistor size has decreased to the point where size can longer be major factor in speed

Multicore processors now fairly common

Increased performance from larger transistor counts and multiple cores has increased energy usage

Texas Tech University

Faster Computers = More Energy

An hour of usage on a super computer today uses the same amount of energy that a moderate home will during the most extreme months of the year

Google estimates their data centers use the same amount of power as 200,000 homes each year.

Texas Tech University

Energy Aware Motivations

Energy Costs

Device Battery Life

Green Computing Initiatives

Texas Tech University

Energy Aware Research

Most work being done in hardware design

CPUs now have multiple operating states to save energy when not in use

Advanced Control Power Interface(ACPI) was developed to give Operating Systems the ability to reduce power consumption of computers

Most models & scheduling techniques rely on altering

CPU operating frequency, which user applications cannot directly access

Texas Tech University

CPU Energy Usage

Energy is the amount of power used for a specified amount of time, 𝐸 = 𝑃 ∗ 𝑑

If the power varies with time then,

𝐸 = 𝑃 𝑑 𝑑𝑑

With N processors, the total energy is the sum of each processor’s usage, 𝐸 =

𝑁 𝑖=0

𝑃 𝑖 𝑑 𝑑𝑑

Texas Tech University

CPU Energy Usage (continued)

The electrical power of a CPU is estimated as 𝑃 = 𝐢𝑉 2 𝐹 ,

𝐢𝑉 2 is a physical constant and F is the operating frequency.

As the frequency of a processor can vary with time, the energy usage of a multicore processor is 𝐸 =

𝑁 𝑖=0

𝐢𝑉

2 𝐹 𝑑 𝑑𝑑

CPUs only operate at S number of frequencies, 𝐸 =

𝑁 𝑖=0

𝑆 𝑗=0

𝐢𝑉 2 𝑓 𝑗 𝑑 𝑖𝑗

Developers cannot select the frequency of the CPU, only if it is idle or not, so there are only 2 frequencies we consider, ON & OFF, 𝐸 =

𝑁 𝑖=0

𝐢𝑉 2 (𝑓

𝑂𝑁 𝑑 𝑖𝑂𝑁

+ 𝑓

𝑂𝐹𝐹 𝑑 𝑖𝑂𝐹𝐹

)

Texas Tech University

Sequential Application Energy

Sequential Applications only use 1 processor, so the other (N-1) processors are idle.

The energy usage is reduced to 𝐸 =

𝐢𝑉 2 (𝑁𝑓

𝑂𝐹𝐹 𝑑

𝑂𝑁

+ 𝑑

𝑂𝐹𝐹

+ 𝑑

𝑂𝑁

(𝑓

𝑂𝑁

− 𝑓

𝑂𝐹𝐹

))

Texas Tech University

Amdahl’s Law

Can be used to give a comparison between sequential & parallel application performance

For this model, it gives us 𝐸

𝑁

, the ratio of the sequential energy usage to the parallel energy usage on an N processor system.

• 𝐸

𝑁

=

𝐢𝑉

2

(𝑁𝑓

𝑂𝐹𝐹

𝑁 𝑖=0 𝑑

𝑂𝑁

+𝑑

𝑂𝐹𝐹

𝐢𝑉 2 (𝑓

𝑂𝑁 𝑑 𝑖𝑂𝑁

+𝑓

𝑂𝐹𝐹

• 𝐢𝑉 2 is constant, so 𝐸

(𝑁𝑓

𝑂𝐹𝐹 𝑑

𝑂𝑁

+𝑑

𝑂𝐹𝐹

+𝑑

𝑂𝑁

𝑁

(𝑓

=

𝑂𝑁

+𝑑

𝑂𝑁

(𝑓

𝑂𝑁

−𝑓

𝑂𝐹𝐹

))

−𝑓

𝑂𝐹𝐹

)) 𝑑 𝑖𝑂𝐹𝐹

)

𝑁 𝑖=0

(𝑓

𝑂𝑁 𝑑 𝑖𝑂𝑁

+𝑓

𝑂𝐹𝐹 𝑑 𝑖𝑂𝐹𝐹

)

Texas Tech University

Observations

Increasing CPU utilization increases Energy

Efficiency

• “Racing to idle” means that the CPU will return to an idle state sooner

Less time executing also means other components will be using less energy too

Texas Tech University

Turning Off Idle Processors

If 𝑓

𝑂𝐹𝐹 is zero, then a parallel application uses the same power as its sequential version

If runtime is fixed, additional processors are unnecessary

Idle CPUs are not turned off and only waste energy

Newer devices have too many CPUs, i.e.

Smart Cell Phones

Texas Tech University

No Idle Power States

If 𝑓

𝑂𝑁

= 𝑓

𝑂𝐹𝐹

, then 𝐸

𝑁

= 𝑆

𝑁

Should only happen if power management settings set incorrectly or poorly

Optimization only way to increase energy efficiency

Download