scx_bpfland: Support balance_power and balance_performance power modes #2016
Conversation
arighi
force-pushed
the
bpfland-auto-power
branch
from
64f9553 to
a790a6f
Compare
Some CPU frequency scaling drivers, such as intel_pstate or amd_pstate, expose intermediate performance scaling preferences: balance_power and balance_performance. These represent a spectrum between maximum power efficiency and peak performance, offering more granular control than the traditional performance or powersave modes. Expose this information in the scx_utils crate, so that schedulers can also apply more fine-grained performance or power-saving optimizations. Signed-off-by: Andrea Righi <arighi@nvidia.com>
Use the following logic when running in auto mode (`-m auto`): +------------------------+----------------+ | Current Power Profile | Active Cores | +------------------------+----------------+ | powersave | slow cores | | balance_power | slow cores | | balance_performance | all cores | | performance | all cores | +------------------------+----------------+ With this logic, on battery-powered systems, the scheduler prioritizes the slow cores when running on battery. When plugged into AC power, it uses all cores more evenly. If the CPU frequency scaling driver supports only a single "balanced" mode, the scheduler operates as balance_performance, using all cores equally. Signed-off-by: Andrea Righi <arighi@nvidia.com>
arighi
force-pushed
the
bpfland-auto-power
branch
from
a790a6f to
f18d94e
Compare
|
After discussing this with the CachyOS community, I decided to slightly change the logic to use all cores also with EPP=performance instead of prioritizing the fast cores. Giving more context on this: auto mode should just try to save power when running on battery or when the powersave EPP is selected, otherwise it should just use all the cores evenly. Prioritizing the fast cores is not always the optimal choice in terms of performance and it can actually lead to sub-optimal performance, especially when the fast cores are slightly faster than the others (e.g., turbo boosted cores). Therefore, it's preferable to let the user decide when to prioritize the fast cores by using |
github-merge-queue
bot
removed this pull request from the merge queue due to failed status checks
CPU frequency scaling drivers, such as intel_pstate or amd_pstate, expose intermediate "balanced" performance scaling preferences:
balance_powerandbalance_performance.Expose them via `scx_utils::autopower, so that scx schedulers can use this information to apply more fine-grained performance / power-saving optimizations.
Adjust the behavior of
scx_bpfland -m autoto use the new power modesbalance_powerandbalance_performance.@multics69 FYI, this includes a small change to scx_lavd as well to support the new
PowerProfile::Balanced { power: true | false }states, which are both mapped toLAVD_PM_BALANCED, so there's no functional change.