Energy Efficiency: From Hardware to Software

Energy efficiency is our contribution to a sustainable future. We not only look at the efficiency of processors but also on how to operate them energy optimal. But what is energy efficiency without measurements?

Data-Dependent Power Consumption

Did you know? The power consumption of an Intel Skylake-SP processor depends not only the C-States, frequencies and used instructions but also on the data processed.

For example, computing an XOR using AVX-512 on all zeros or all ones makes a difference of 58 W. [1]

Varying power consumption depending on the data handed to AVX XOR.
Vampir visualization of data-dependent power at nominal frequency. Over time, the bit count of value1 and value2 is changed (top and middle panel). The full system AC power consumption shown is in the bottom panel.

READEX: Optimizing the Energy Consumption of your HPC Application

Did you know? The energy consumption of an application can be reduced without sacrificing too much performance.

Simply adapt the core- and uncore frequency of the processor to your application’s needs. For the Kripke mini-app, this leads to a reduction of the energy consumption up to 15% [2].

Moreover, there is no need to find the best frequency yourself. Thanks to the precision of Intels RAPL [3], this can be done automatically using reinforcement learning [2].

Varying energy savings of the Kripke benchmark depending on the amount of nodes
The plot shows the Kripke mini-app's potential energy savings, depending on the number of nodes used. The values noted at the grey line indicate the runtime, while the line itself shows the runtime normalized to the default runtime. The superlinear speedup from 1 node to 2 nodes might be related to MPI activation and a resulting better load balancing.
The picture shows a function from Kripke, which is optimized for energy consumption. The red arrow indicates the overall optimization direction. A minimum is reached in less than 50 steps. While the color shows the iterations spend with a particular configuration, numbers indicate the last measured energy consumption at this specific configuration.

Metric-Q: A Scalable Infrastructure for your High-Resolution Time Series Data

Did you know? You can access your time series data with O(1) instead of O(N) without sacrificing any information.

While MetricQ helps you to collect, analyze and save your sensor data fast, the Hierarchical Timeline Aggregation (HTA) requires moderate data processing during data collection while allowing to reduce the complexity of a typical timeline request.

Using this advanced technique, you can visualize billions of data points, e.g., from your power measurements, in an instant [4].


[1] Robert Schöne, Thomas Ilsche, Mario Bielert, Andreas Gocht and Daniel Hackenberg:
Energy Efficiency Features of the Intel Skylake-SP Processor and Their Impact on Performance.
In: 2019 International Conference on High Performance Computing & Simulation (HPCS)
DOI: 10.1109/HPCS48598.2019.9188239 [PDF]

[2] Andreas Gocht, Robert Schöne and Mario Bielert:
Q-Learning Inspired Self-Tuning for Energy Efficiency in HPC
2019 International Conference on High Performance Computing & Simulation (HPCS)
DOI: 10.1109/HPCS48598.2019.9188112

[3] Daniel Hackenberg, Robert Schöne, Thomas Ilsche, Daniel Molka, Joseph Schuchart and Robin Geyer:
An Energy Efficiency Feature Survey of the Intel Haswell Processor
2015 IEEE International Parallel and Distributed Processing Symposium Workshop
DOI: 10.1109/IPDPSW.2015.70 [PDF]

[4] Thomas Ilsche, Daniel Hackenberg, Robert Schöne, Mario Bielert, Franz Höpfner and Wolfgang E. Nagel:
MetricQ: A Scalable Infrastructure for Processing High-Resolution Time Series Data
2019 IEEE/ACM Industry/University Joint International Workshop on Data-center Automation, Analytics, and Control (DAAC)
DOI 10.1109/DAAC49578.2019.00007 [PDF]

More Information

Please send us a message if you have any questions!