Linux Plumbers Conf 2010

One Line Summary

Abstract

DVFS is a technique that uses the optimal operating frequency and voltage to allow a task to be performed in the required amount of time. Traditionally DVFS has been associated with the processor speed scaling or the bus speed scaling along with scaling the associated voltage regulator. In the new approach we would like to move away from the concept of just processor speed or bus speed and extend DVFS to the entire devices present in the system. A system can have multiple scalable devices each one of them associated with a voltage regulator. Depending on the voltages to which the regulator can be scaled, the device can have varied operating frequencies (scalable frequencies). This implies each scalable device can have a series of voltage and operating frequency pair associated with it. Also a voltage regulator can be shared between devices (scalable/ non-scalable). This means that the final voltage to which the regulator is to be scaled at any instant should be the maximum of the operating voltage requirements of all the devices fed by the regulator. If different VDDs have any inter dependencies, that also gets taken care in this approach by having a cross vdd dependency table for each VDD. Also as frequency of a device for a given Voltage is actually the maximum frequency for the given voltage, it is possible to restirct the frequency to a lower one depending on the thermal condition (Thermal management).

The design involves maintaining a table per scalable device capturing the information like the voltage regulator associated with the device, a table containing voltage-operating frequency pairs, function pointer to populate the actual function that scales the device operating frequency etc. Also a use counting mechanism is needed on a per regulator basis to take in requests from various devices and return back the maximum of all the requests. A generic kernel level API can then take in the device pointer and the required rate as parameters and derive the actual rate possible after taking in the possible voltage from the voltage regulator use-counting mechanism. This API will call into the function pointer populated in the per scalable device table for scaling the operating frequency of the device and a voltage scaling API to scale the associated
regulator voltage.

The idea is currently undergoing implementation for TI OMAP processors and the presentation/paper elucidates how this is being achieved in OMAP and how this principle can be extended to other Soc’s.

This abstract is based on Collective work of Thara Gopinath (thara@ti.com) and Benoit Cousson (b-cousson@ti.com) with inputs from Kevin Hilman, Paul Walmsley, Anand Sawant, Partha Basak and Vishwanath BS.

Speaker

• Thara Gopinath

  Biography

  Thara is working as a Senior Software in Texas Instruments India. She has extensively worked on TI OMAP Proecessors in the area of Device Drivers and specialized in OMAP Power Management with various versions of Linux Kernel.