1. Progress Report
2012/12/20

2. Computation Offloading
Mobile devices have limited energy and computing resources.
Offloading some workloads to remote servers leads to:
Power-saving.
Performance improving.
Shorter execution time.
Better results.

3. Our Idea
Instead of studying offloading policies, we aim at the effects caused by offloading.
After offloading, a computation-intensive process becomes I/O-intensive.
Does this phenomenon affects scheduler?
Does this phenomenon affects DVFS?
Does this phenomenon affects cache/memory?
…etc.

4. Problem Service thread Computation-intensive => I/O-intensive
Task in waiting state will not be scheduled.

5. Current Flow
Offloading Framework
Scheduler
DVFS
DPM
Load change

6. New strategy The original idea of virtual core is:
N cores
C-task
Remote
cores
Computation Offloading
N+1 cores

7. New strategy(Cont.) 反其道而行 Close a core after offloading
N cores
C-task
Remote
cores
Computation Offloading
N-1 cores

8. Reason
After offloading, the rest of the tasks will be scheduled to N cores.
Should have better performance.
Does not guarantee energy saving.
All the cores are still working!
If we close a core after offloading
Imagine that the computation-intensive task is non-preemptive, but consume zero power.
Energy saving with little effect to (other tasks) performance.

9. New Flow
Offloading Framework
Scheduler
DVFS
DPM

10. Performance of other tasks
Comparison
Performance of other tasks
Power Consumption
No offloading --
Highest
Virtual core (N+1)
Better than No offloading
Low
Proposed strategy (N-1)
Similar to No offloading
Lowest

11. However This is a theoretically strategy.
Need to design some experiments to verify the strategy.

12. Possible Topics Close more cores Close which core(s)
-1?-2?-n/2?-(n-1)?
Close which core(s)
The one executing the offloaded task?
Cache related
Theoretical problem
Math model