Messages in this thread | | | Date | Wed, 26 Aug 2009 16:37:21 +0530 | From | Arun R Bharadwaj <> | Subject | [v2 PATCH 0/2]: cpuidle: Introducing cpuidle infrastructure to POWER |
| |
Hi,
"Cpuidle" is a CPU Power Management infrastrusture which helps manage idle CPUs in a clean and efficient manner. The architecture can register its driver (in this case, pseries_idle driver) so that it subscribes for cpuidle feature. Cpuidle has a set of governors (ladder and menu), which will decide the best idle state to be chosen for the current situation, based on heuristics, and calculates the expected residency time for the current idle state. So based on this, the cpu is put into the right idle state.
Currently, cpuidle infrasture is exploited by ACPI to choose between the available ACPI C-states. This patch-set is aimed at enabling cpuidle for powerpc and provides a sample implementation for pseries.
Currently, in the pseries_dedicated_idle_sleep(), the processor would poll for a time period, which is called the snooze, and only then it is ceded, which would put the processor in nap state. Cpuidle aims at separating this into 2 different idle states. Based on the expected residency time predicted by the cpuidle governor, the idle state is chosen directly. So, choosing to enter the nap state directly based on the decision made by cpuidle would avoid unnecessary snoozing before entering nap.
This patch-set tries to achieve the above objective by introducing a pseries processor idle driver called pseries_idle_driver in arch/powerpc/platform/pseries/processor_idle.c, which implements the idle loop which would replace the pseries_dedicated_idle_sleep() when cpuidle is enabled.
Experiment conducted: ----------------------
The following experiment was conducted on a completely idle JS22 blade, to prove that using cpuidle infrastructure, the amount of nap time increases.
Nap and snooze times were sampled for all the cpus. For a window of 1000 samples, When cpuidle was enabled, the total nap time was of the order of a few seconds (5-10s), whereas the total snooze time was of the order of a few milliseconds(10-30 ms).
When cpuidle infrastructure was disabled and the regular pseries_dedicated_idle_sleep() idle loop was used, the snooze time itself was of the order of hundreds of milliseconds. (100 - 500 ms). This is clearly due to unnecessary snoozing before napping even on a completely idle system.
The previous post in this area can be found at http://lkml.org/lkml/2009/8/19/150
Changes from the previous version: ----------------------------------
Earlier I used the the name TPMD (Thermal and Power Management Devices) to refer to this module which hooks on to the cpuidle infrastructure. I have renamed this to Pseries Processor Idle, in order to avoid confusion.
Patches included in this set: ------------------------------ PATCH 1/2 - Enable cpuidle for pSeries. PATCH 2/2 - Implement Pseries Processor Idle idle module
Any feedback on the overall design and idea is immensely valuable.
--arun
| |