The PowerClamp driver

Most users will never want to use PowerClamp. As a general rule, when one has purchased hardware with a given computational capability, one wants that full capability to be available when needed. But there are situations where it makes sense to run a system below its full speed. Data centers have power-consumption and cooling constraints that can argue against running all systems flat-out all the time. Even the owner of an individual laptop or handheld system may wish to ensure that its operating temperature does not exceed a given value; an overly hot laptop can be uncomfortable to work with, even if it is still working within its specified temperature range. So there can be value in telling the system to run slower at times.

The PowerClamp driver allows the system administrator to set a desired idle percentage by way of a sysfs attribute. That percentage is capped at 50% in the current implementation. Once a percentage has been set, the kernel monitors the actual idle time for each processor in the system. Should a processor's idle time fall below the desired idle percentage, a special kernel thread (called kidle_inject/N, where N is the number of the CPU to which the thread is assigned) is created to take corrective action.

That thread operates as a high-priority realtime process, so it is able to respond quickly when needed. Its job is relatively simple: look at the amount of idle time on its assigned CPU and calculate the difference from the desired idle time. Then, periodically, the thread will run, disable the clock tick, and force the CPU into a sleep state for the required amount of time. The sleeping is done for a given number of jiffies, so the sleep states tend to be relatively long — a necessary condition for an effective reduction in power usage.

Naturally, the PowerClamp thread will continue to monitor actual idle time as it operates, adjusting the amount of forced sleep time as needed. It also monitors the amount of desired sleep time that is lost to interrupts. Interrupts remain enabled during the forced sleep, so they can bring the processor back to an operational state before the PowerClamp driver would have otherwise done so. Over time, the amount of sleep time lost in this manner is tracked; the driver will then attempt to compensate by increasing the amount of forced sleep time to try to pull the CPU back to the original idle time target.

By itself, PowerClamp can come close to achieving the desired level of idle time on a system with a changing workload. Often, though, the real goal is not idle time as such; instead, the purpose is to keep the system within a given level of power consumption or a set of thermal limits. Doing that will require the implementation of additional logic in user space. By monitoring the parameter of interest, a user-space process can implement a control loop that adjusts the desired level of idle time as needed. The PowerClamp driver can respond relatively quickly to those changes, giving the control process an effective tool for the management of the amount of power used by the system.

The driver has been through a couple of revisions with little in the way of substantive comments. This patch poses a relatively small risk to the system, since it does not do anything if the feature is not in use. It could thus conceivably be ready for merging as soon as the 3.8 development cycle. Some more information can be found in the documentation file included with the patch.