Communications of the ACM - September 2018

tion logs are captured. In this example, time-domain analysis is not practical for gathering scalability and utilization. Instead, a histogram analysis identifies in which bucket the maximum number of load points lie for a given frequency of operation. The peak of the histogram represents the load value with the highest probability of occurrence for the given frequency. In the example data in Figure 5, about 500 samples (or 80% of total samples) occurred at a 33% load. Such inferences are obviously hard to derive from the time chart in Figure 4 even if the frequency is fixed.

The experiment is iterated through all other frequencies, one at a time, to get a range of peak histogram values, as shown in Figure 6.

With the same data from the set of experiments, a similar histogram ridge trace for scale-factor can be plotted as shown in Figure 7.

Scaling of scale factor. A plotting
tool is used to create a map-view of
peak histogram points (in scale factor)
in Figure 7, which is then plotted with
red dots in Figure 8. On the same map-
view, the scaling of scale factor equa-
tion is plotted for different possible
target frequencies based on any one
initial seed value for utilization and
scale factor. The correlation between
the equation and the full data set is
shown in Figure 8.

Scaling of load. The map-view of peak histogram points of utilization (Figure 6) is plotted with red dots in Figure 9. On the same map view, the scaling of load equation is plotted for different possible target frequencies based on one initial seed value for utilization and scale factor.

As summarized by the results shown in these figures, the independent estimates of the equation fit nearly precisely with the actual experiment’s statistical results. This methodology is applied to multiple other workloads (with different scale factors and load levels) and verified as noted here.

Conclusion

This article elaborated on the causali-
ties of utilization and scale factor in
relation to possible frequency-selec-
tion choices within a time window.
Based on this, generic workload scal-
ing equations are derived, quantify-
ing utilization impact. The equations
are verified by applying the histogram
ridge trace method at discrete DVFS
block level. Though only the CPU
core example was detailed, similar
equation agreements were observed
on other DVFS blocks (graphics sub
blocks), other workloads and other
operating systems (Windows/Linux).
The equation-estimated curve corre-
lates very accurately with the actual
ridge trace curve. This implies the uti-
lization impact can be “predicted” to
be statistically accurate in every cycle
and any divergence from it can be at-
tributed to workload-inherent utiliza-
tion change in that cycle and treated
as appropriate to the solution using it.

Acknowledgments

The author is thankful for all the help from colleagues at Intel, specifically, valuable input from principal engineers Harinarayanan Seshadri and Rajeev Muralidhar and software engineer B. M. Shravan K.

Related articles on queue.acm.org

Power-Efficient Software

Eric Saxe https://queue.acm.org/detail.cfm?id=1698225

Maximizing Power Efficiency with Asymmetric Multicore Systems

Alexandra Fedorova et al. https://queue.acm.org/detail.cfm?id=1658422 Energy Management on Handheld Devices Marc A. Viredaz et al. https://queue.acm.org/detail.cfm?id=957768

References

1. Howse, B. Examining Intel’s new Speed Shift tech on Skylake: more responsive processors. Anand Tech, 2015; https://bit.ly/2Jf3Gxq.

2. Intel. Intel 64 and IA- 32 Architectures, Software Developer’s Manual, Volume 3B: System Programming Guide, Part 2; https://intel.ly/2sKeSqA

3. Kidd, T. Power management states: P-states, C-states, and Package C-states. Intel Developer Zone, 2014; https://intel.ly/2xM094b

4. Power Shaping and Stress Tool (PSST) for Intel Platforms. https://github.com/intel/psst

Noor Mubeen is a software architect at Intel client R&D group, focusing on power, thermal, and energy management breakthroughs. He has over 17 years’ experience spanning a breadth of domains including networking, storage file systems, and embedded devices.