DAMON-Based Memory Reclamation Merged For Linux 5.16

[phoronix]

Phoronix: DAMON-Based Memory Reclamation Merged For Linux 5.16

Following Amazon's DAMON being merged in Linux 5.15 as a data monitoring access framework, being merged for Linux 5.16 is an addition building on top of that for memory reclamation when experiencing system RAM pressure...

https://www.phoronix.com/scan.php?page=news_item&px=DAMON-Reclamation-Linux-5.16

[tildearrow]

At last! A kernel solution to the high memory usage slowdown...

...but it's only the first part.

[Mitch]

This is so awesome. For low-end devices and for servers, this has to be a life-saver. I imagine the one edge-case where it wouldn't be so good would be an older laptop where that kind of CPU would measurably hurt battery. As big-little becomes more widespread and better though, background stuff like that should be negligible on power consumption.

[jonix]

Reclaiming memory before it turn into a problem seems to be the right move.

As big-little becomes more widespread and better though, background stuff like that should be negligible on power consumption.

I am counting on that as well.

[geearf]

Is this expected to be useful on the desktop as well? Maybe I can finally get excited about a new kernel release (well there's already the zstd update).

[yump]

5% CPU is not something I would call "only" on a battery powered machine, but it seems to pretty much directly correspond to the 50 ms/s quota. I wonder how it interacts with the CPU governor? If the quota is real-time, the number of cycles it gets would scale with the current frequency, so if the reclamation itself can be kept from triggering the governor to increase the frequency, 5% might not be too awful. But uclamp is an "Android thing", and maybe not even hooked up on intel_pstate HWP.

The fragmented approaches to problems in the kernel is so annoying. Based on the benchmarks, it looks like Intel managed to release Alder Lake without properly hooking it up to capacity-aware scheduling, because that's an "Arm thing". And this DAMON seems to be tackling a very similar problem space as multi-generational LRU, except that's Google and this is Facebook.