JakeHillion/scx-upstream
1
0
Fork 0
mirror of https://github.com/sched-ext/scx.git synced 2024-11-28 13:40:28 +00:00
Code Issues Packages Projects Releases Wiki Activity

scx_layered: Add more docs on tuning

Browse Source 
Signed-off-by: Daniel Hodges <hodges.daniel.scott@gmail.com>
This commit is contained in:
Daniel Hodges 2024-08-28 12:37:46 -07:00
parent 96326b1ef3
commit a23308e7b0
1 changed files with 19 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
scheds/rust/scx_layered/README.md
View File

@ -49,4 +49,22 @@ the `perf` field can be set. This must be used in combination with the
`schedutil` frequency governor. The value should be from 0-1024 with 1024 being
maximum performance. Depending on the system hardware it will translate to
frequency, which can also trigger turbo boosting if the value is high enough
and turboe is enabled.
and turbo is enabled.
Layer affinities can be defined using the `nodes` or `llcs` layer configs. This
allows for restricting a layer to a NUMA node or LLC. Layers will by default
attempt to grow within the same NUMA node, however this may change to suppport
different layer growth strategies in the future. When tuning the `util_range`
for a layer there should be some consideration for how the layer should grow.
For example, if the `util_range` lower bound is too high, it may lead to the
layer shrinking excessively. This could be ideal for core compaction strategies
for a layer, but may poorly utilize hardware, especially in low system
utilization. The upper bound of the `util_range` controls how the layer grows,
if set too aggressively the layer could grow fast and prevent other layers from
utilizing CPUs. Lastly, the `slice_us` can be used to tune the timeslice
per layer. This is useful if a layer has more latency sensitive tasks, where
timeslices should be shorter. Conversely if a layer is largely CPU bound with
less concerns of latency it may be useful to increase the `slice_us` parameter.
`scx_layered` can provide performance wins, for certain workloads when
sufficient tuning on the layer config.