Running with CONFIG_HZ_PERIODIC=y keeps the scheduler tick running
continuously, which produces higher jitter and lower power efficiency.
In contrast, CONFIG_NO_HZ_IDLE=y (the upstream default) stops the tick
only when the CPU is idle, giving lower idle power and normal runtime
jitter.
An Intel N150-based router/firewall was tested using two kernel builds:
one with CONFIG_HZ_PERIODIC=y, and one with CONFIG_NO_HZ_IDLE=y. Power
consumption was measured while the system was essentially idle (no
meaningful traffic). The CONFIG_NO_HZ_IDLE=y build consistently used
less power.
Details: The two power-measurement methods were:
1. PkgWatt from turbostat (software)
2. Wall-power measurement using a Kill-A-Watt (hardware)
The test began by zeroing the Kill-A-Watt and simultaneously running:
turbostat --quiet --Summary --interval 10 --show Busy%,PkgWatt
The test duration was defined by the time required for the Kill-A-Watt
to accumulate 0.005 kWh, after which the average wattage was calculated.
Results:
+----------------------+-----------------+----------------+-----------+
| Metric | HZ_PERIODIC | NO_HZ_IDLE | Delta % |
+----------------------+-----------------+----------------+-----------+
| PkgWatt | 3.59 ± 0.38 | 3.38 ± 0.34 | -5.9 % |
+----------------------+-----------------+----------------+-----------+
| Avg wattage at wall | 12.47 | 12.00 | -3.77 % |
+----------------------+-----------------+----------------+-----------+
The mean PkgWatt difference is 210 mW (5.9%) in favor of
CONFIG_NO_HZ_IDLE=y, with a t-statistic of ~3.17 and p ≈ 0.002.
Wall-power measurements show a 470 mW (3.77%) reduction under
CONFIG_NO_HZ_IDLE=y.
Signed-off-by: John Audia <therealgraysky@proton.me>
Link: https://github.com/openwrt/openwrt/pull/21470
Signed-off-by: Robert Marko <robimarko@gmail.com>