Run-time Power Estimation for Mobile and Embedded ...
Transcript of Run-time Power Estimation for Mobile and Embedded ...
idle cstm int cstm fp basicmath bitcount qsort susan typeset dijkstra patricia ispell adpcm FFT h264 hq h264 lq mpeg4 hq mpeg4 lq 40 unseenWorkloads
0.00
0.05
0.10
0.15
0.20
Pow
er(W
)
1.56%
6.66% 7.25%
7.18%
1.57%
6.85%
1.95% 7.09% 4.45%2.66%
3.30%
1.17%
2.86%2.57%
3.26%
4.37% 4.01%
5.05%MeasuredEstimated
Run-time Power Estimation for Mobile and Embedded Asymmetric Multi-Core CPUs
Matthew Walker, Dr Geoff Merrett, Professor Bashir Al-Hashimi{mw9g09, gvm, bmah}@ecs.soton.ac.uk
Electronic and Software Systems Research Group, Electronics and Computer ScienceUniversity of Southampton
Performance Counter (PMC) Power Model
• Performance counter events (e.g. L2 cache miss, branch mis-prediction) correlate well with power consumption;
• A run-time power model was built for a BeagleBoard-xM (Figure 3)
• Extremely accurate: <3.2% error across large range of workloads when running in real-time (Figure 2)
Conclusion• Two run-time power models built; PMC-based model
more accurate but less practical than utilisation-based model
• Utilisation model can an predict the power profile of one core from statistics from another in a big.LITTLE system
Further Work• Implement on Android and test on real smartphone• Analyse big.LITTLE trade-offs - how to make the
smartest decisions • Use to aid run-time management
Utilisation Power Model• Problem with PMCs: they are difficult/impossible to
obtain on most mobile/embedded devices• Will a simpler metric do?• Power model using simple CPU utilisation was built on a
Samsung big.LITTLE SoC (used in Samsung Galaxy S5, Chromebook 2, Samsung Galaxy Note 3 - all released in 2014)
• Error of 5.6% on 'little' Cortex-A7 and 7.2% on 'big' Cortex-A15 (per-core power estimation)
• Utilisation models can be applied to any platform• Power of each task can be estimated• Can foresee how much power a task would consume if it
were running with a different core/frequency (Error: 10%)
Introduction• A run-time manager (RTM) can make significant energy
savings by making smart decisions when controlling the processor's operation (e.g. DVFS, DPM, task-core mapping)
• To make smart decisions, it needs to know (in real-time) how much power is currently being consumed
• Aim of this research is to built run-time power models
Workload Find Correlation Model
Stats
Power
Figure 1 Simplified experiment methodology
Figure 3 BeagleBoard-xM Figure 4 ODROID-XU+E Board
Development Platforms
www.prime-project.org
Figure 2 Run-time power and estimated power from PMC model
Figure 4 Utilisation model power and predicted power across workloads