Minimizing%20energy%20consumption%20for%20pipelined%20real%20time%20applications

1
Minimizing energy consumption for pipelined real
time applications

• Ahmed Abousamra

2
Problem

• Periodic job with interarrival time T, and
  response time DxT, where x 1,2,
• Job is composed of tasks that are executed in a
  pipeline. WCEC c

3
Solution for WCEC

• Power function for a core (pipeline stage)
• Because DxT, we can have x pipeline stages
• But because of static and dynamic power
  components, optimal number of stages (assuming
  divisible load) is

4
Solution for WCEC

• In reality, load is not divisible to any number
  of fractions.
• Operating frequency, f, can only be changed in
  steps.
• Solution components
• Number of stages
• Mapping of tasks to stages
• Frequency of each stage

5
What if a job needs less than WCEC?

• Reclaim slack time to consume less energy

T 10 seconds
Jm
Jm1
Jm2
6 seconds
8 seconds
At time 8
Jm2
Jm1
Can execute for 12 seconds thus can reduce
frequency and save energy
6
What if a job needs less than WCEC?

• Reclaim slack time tell previous stage(s) to
  speed up

T 10 seconds
Jm
Jm1
Jm2
6 seconds
speedup
At time 8
Jm1
Jm2
Both stages can execute for 12 seconds thus can
reduce frequency and save energy
7
What if a job needs less than WCEC?

• Goal Minimize expected energy consumption
• Use probability distribution of completing each
  task in less than a certain number of cycles to
  map tasks to stages.
• At each stage, start with a low speed, then
  gradually increase it as the deadline of the
  stage approaches.
• Can also reclaim slack time

8
Questions?