scx_utils, scx_lavd: Add cpu_capacity to Topology and use it in scx_lavd. #1465
Conversation
|
Ooops, it fails in our CI. I will take a look. |
multics69
force-pushed
the
lavd-cpu-capacity
branch
from
ddf6320 to
e8d4260
Compare
Try to make the best guess of CPU capacity by reading various sysfs sources from the most precise to the least precise in the following order: /sys/devices/system/cpu/cpuX/cpufreq/amd_pstate_highest_perf /sys/devices/system/cpu/cpuX/acpi_cppc/highest_perf /sys/devices/system/cpu/cpuX/cpu_capacity /sys/devices/system/cpu/cpuX/cpufreq/cpuinfo_max_freq The capacity value of each CPU is scaled to 1024. Signed-off-by: Changwoo Min <[email protected]>
Instead of guessing CPU capacity using a CPU's max frequency, rely on cpu_capacity information from Topology, which is best guessed from various sysfs sources. Signed-off-by: Changwoo Min <[email protected]>
multics69
force-pushed
the
lavd-cpu-capacity
branch
from
e8d4260 to
3ebc165
Compare
Fixed. |
There was a problem hiding this comment.
LGTM. Few comments.
The static approach works in most of the cases, but in some architectures the capacity can change at runtime (see for example #1207). In general it'd be nice to have an event-based API to catch topology changes. But I'm not sure how to implement something like this.
Or maybe just have an API similar to scx_bpf_cpuperf_cap() to get a more accurate capacity value, see for example get_actual_cpu_capacity() in fair.c, that takes in account the cpufreq pressure, thermal throttling and potentially other factors.
But this is not related to this PR, just some thoughts for the future.
|
That's definitely a way to go. Thanks for the suggestion. |
github-merge-queue
bot
removed this pull request from the merge queue due to failed status checks
github-merge-queue
bot
removed this pull request from the merge queue due to failed status checks
There are various sysfs sources that somehow represent each CPU's capacity even if the kernel does not properly implement arch_scale_cpu_capacity(), which is used for scx_bpf_cpuperf_cap().
This PR tries to make a best guess about CPU capacity based on available sysfs entries by checking the most precise (e.g., vendor-specific information) to the least precise (CPU max frequency).
Also, scx_lavd is changed to use cpu_capacity exposed by Topology.