Carsten Emde, Open Source Automation Development Lab

When the first low-power processors appeared and initial real-time tests were run on them, it quickly became evident that the goals of achieving real time and power saving are incompatible. As a result, the straight-forward recommendation was to completely disable any power-saving mechanism, if real time is required. The question is then: What can you do, if the real-time device runs on small batteries or has thermal constraints and thus relies on power saving?

In order to better study the relationship between power-saving features, real-time capabilities, and total system power consumption, several Linux systems in our test center were equipped with continuous power metering.

In addition, the kernels of these systems were upgraded with a backport of a recently available kernel fix that makes it possible to enable the various sleep states individually per processor core. Recordings of sleep states, clock frequency per processor core, and total system power consumption were added to the already existing long-term data recording with a temporal resolution of five minutes.

In single-core systems, the hypothesis was tested that power saving may in fact result in longer latencies but does not counteract real-time determinism in general.

In multi-core systems, the hypothesis was tested as to whether it is possible to disable throttling and sleep states of one of the cores, run a real-time application with low latencies on it, and still enjoy a reduced power consumption when the other cores are idle or running under a low load.

Linux 3.2 or 3.4 with PREEMPT_RT real-time was used in all cases.

The results clearly demonstrate that real-time capabilities and power saving may coexist. As a penalty, however, enabling power saving results in longer latencies. When throttling is allowed, the increase of the latencies is approximately proportional to the decrease in the clock frequency. When sleep states are enabled, latencies may increase drastically, but this increase can be limited, if deeper sleep states are disabled. Allowing the processor to enter only a light sleep state may already yield measurable power savings.

In multi-core systems, it is possible to restrict the real-time process to one of the cores and to disable all power saving in this core. This will result in perfect real-time capabilities for this core while the other cores may still run at low clock frequencies and enter deep sleep states; consequently, they do not consume more energy than necessary and thus may contribute to reduced power consumption for the entire system.