Allow to sleep instead of busy waiting when limiting refresh rate #1563
Conversation
On single core boards, busy waiting renders the system unresponsive.
We add the --led-no-busy-waiting flag, to allow the process to sleep
instead of busy waiting, thereby freeing up CPU cycles for other
processes at the cost of slightly less accurate frame timing.
On a Raspberry Pi 3B on two 32x32 panels, busy waiting keeps a core
98% busy. Sleeping reduces it to 14%.
On a Raspberry Zero on two 32x32 panels, busy waiting keeps the single
core 95% busy and gives an unresponsive system. Sleeping reduces it to
32% and keeps the system responsive.
Sleeping degrades the timing accuracy a bit, but is still good enough
by a far margin for many uses. The following numbers are for driving
two 32x32 panels and averaging the refresh rate across a minute (the
1kHz refresh rate limit is of course not reached, and is only there to
make sure that the relevant code path is taken):
+--------------+-----------+
| busy waiting | sleeping |
+--------------------------------------------------+--------------+-----------+
| Raspberry Pi 3B, limit 70Hz, hardw. pulsing | 70.000Hz | 69.977Hz |
| Raspberry Pi 3B, limit 70Hz, no hardw. pulsing | 70.000Hz | 69.977Hz |
| Raspberry Pi 3B, limit 1kHz, hardw. pulsing | 430.824Hz | 430.855Hz |
| Raspberry Pi 3B, limit 1kHz, no hardw. pulsing | 365.498Hz | 365.456Hz |
| Raspberry Pi Zero, limit 70Hz, hardw. pulsing | 69.986Hz | 69.643Hz |
| Raspberry Pi Zero, limit 70Hz, no hardw. pulsing | 69.986Hz | 69.638Hz |
| Raspberry Pi Zero, limit 1kHz, hardw. pulsing | 263.877Hz | 264.538Hz |
| Raspberry Pi Zero, limit 1kHz, no hardw. pulsing | 233.166Hz | 231.650Hz |
+--------------------------------------------------+--------------+-----------+
|
Hi @somechris, thanks for this improvement, this really helps lowering the CPU load, very useful on a Pi Zero :) |
|
Nice! Thanks @somechris The code looks great. Thanks for taking care that this is an inverted option so that memsetting the struct to zero would be backward compatible. So code-wise, I think we're ready to merge, but you should also add the description of this flag in a paragraph in the README where |
GitHub's Markdown renderer underlined the whole text from `--led-no-drop-privs` until the start of the next section, which made it hard to read. Be converting `A` tag from self-closing to explicitly closed tag, GitHub renders the file without the extra underlining again, which makes it easier to read.
Done :-) |
On single core boards, busy waiting renders the system unresponsive. We add the --led-no-busy-waiting flag, to allow the process to sleep instead of busy waiting, thereby freeing up CPU cycles for other processes at the cost of slightly less accurate frame timing.
On a Raspberry Pi 3B on two 32x32 panels, busy waiting keeps a core 98% busy. Sleeping reduces it to 14%.
On a Raspberry Zero on two 32x32 panels, busy waiting keeps the single core 95% busy and gives an unresponsive system. Sleeping reduces it to 32% and keeps the system responsive.
Sleeping degrades the timing accuracy a bit, but is still good enough by a far margin for many uses. The following numbers are for driving two 32x32 panels and averaging the refresh rate across a minute (the 1kHz refresh rate limit is of course not reached, and is only there to make sure that the relevant code path is taken):
+--------------------------------------------------+--------------+-----------+
| Raspberry Pi 3B, limit 70Hz, hardw. pulsing | 70.000Hz | 69.977Hz |
| Raspberry Pi 3B, limit 70Hz, no hardw. pulsing | 70.000Hz | 69.977Hz |
| Raspberry Pi 3B, limit 1kHz, hardw. pulsing | 430.824Hz | 430.855Hz |
| Raspberry Pi 3B, limit 1kHz, no hardw. pulsing | 365.498Hz | 365.456Hz |
| Raspberry Pi Zero, limit 70Hz, hardw. pulsing | 69.986Hz | 69.643Hz |
| Raspberry Pi Zero, limit 70Hz, no hardw. pulsing | 69.986Hz | 69.638Hz |
| Raspberry Pi Zero, limit 1kHz, hardw. pulsing | 263.877Hz | 264.538Hz |
| Raspberry Pi Zero, limit 1kHz, no hardw. pulsing | 233.166Hz | 231.650Hz |
+--------------------------------------------------+--------------+-----------+