Google Proposes Multi-Generational LRU For Linux To Yield Much Better Performance

[phoronix]

Phoronix: Google Proposes Multi-Generational LRU For Linux To Yield Much Better Performance

Google engineer Yu Zhao sent out patches proposing a "multigenerational LRU" implementation for the Linux kernel's least recently used (LRU) handling for memory page replacement...

https://www.phoronix.com/scan.php?page=news_item&px=Linux-Multigen-LRU

[tildearrow]

Linux was the fastest, and now it gets even faster? Amazing!

[Jumbotron]

Just when Google pisses me off enough to go through the "Reality Distortion Field" to enter into the walled garden that is Apple...their engineers pull off something like this that benefits that wider Linux community. Ok Google....you have me for a little bit longer. Keep doing this and you'll have me longer still. But if you keep up with the killing off of good apps and completely FUBARing existing good apps like the old Google Wallet which is now the complete steaming dog turd known as Google Pay, I may yet trip the Apple Fantastic.

[microcode]

Very cool. Always room in my heart for better algorithms that are simpler to implement than inferior ones; and I will have to look at this paper for work now, since one of the problems we have is very similar to this problem.

[kapouer]

I suppose i'm the only one reading "ADAPTATIVE REPLACEMENT CACHE" here.
(Which has a patent granted to IBM).

[CommunityMember]

I suppose i'm the only one reading "ADAPTATIVE REPLACEMENT CACHE" here.
(Which has a patent granted to IBM).

As IBM is a member of OIN, usage of that patent (should it even apply to this proposal) should not be a problem.

[muncrief]

I'm running 5.12-rc2 with the LRU patches right now and nothing is obviously broken. What PTS tests should I run to best view performance improvements? I have an rc2 without the patches I want to compare side by side.

[waxhead]

I'm running 5.12-rc2 with the LRU patches right now and nothing is obviously broken. What PTS tests should I run to best view performance improvements? I have an rc2 without the patches I want to compare side by side.

While you're on that , would you mind exploring the effects of /proc/sys/vm/swappiness as well? it would be interesting to see how this performs with swapiness values set to the extreme such as 10 and 100 or even 200 (which became possible in kernel 5.8). I am curious how the system performs while recovering from being idle for some time especially since the ageing calculation is done differently.

[yoshi314]

Just when Google pisses me off enough to go through the "Reality Distortion Field" to enter into the walled garden that is Apple...their engineers pull off something like this that benefits that wider Linux community. Ok Google....you have me for a little bit longer. Keep doing this and you'll have me longer still. But if you keep up with the killing off of good apps and completely FUBARing existing good apps like the old Google Wallet which is now the complete steaming dog turd known as Google Pay, I may yet trip the Apple Fantastic.

i think your problem is believing that there is only Apple and Google out there.

[MastaG]

It would be nice if someone could post backports of this patch sets for previous kernels. So we can try them on sbc's which tend to be tight on ram.