A New Governor Continues To Be Worked On For Linux Kernel's CPUFreq

[phoronix]

Phoronix: A New Governor Continues To Be Worked On For Linux Kernel's CPUFreq

The Linux kernel's ACPI and power management subsystem maintainer, Rafael Wysocki, today sent out the latest patches on a new governor for the CPUFreq CPU frequency scaling driver for the Linux kernel...

http://www.phoronix.com/scan.php?page=news_item&px=CPUFreq-schedutil-Nearly-Ready

[uid313]

Do people still have problems with the mouse pointer freezing during load?
Or audio skipping?

[s.ivanov]

Unfortunately I'm still unable to use my desktop while the system is disk busy BFQ helps to an extend but not much setting IONICE levels make no impact on the problem

[FireBurn]

Do people still have problems with the mouse pointer freezing during load?
Or audio skipping?

I've not seen that in a while on either my performance Skylake system or my low powered Kabini - everything works fine with things like LLVM & Mesa compiling in the background

[s.ivanov]

everything works fine with things like LLVM & Mesa compiling in the background

SSD or spinning drives?

[FireBurn]

They're both running with SSDs

[sarfarazahmad]

Unfortunately I'm still unable to use my desktop while the system is disk busy BFQ helps to an extend but not much setting IONICE levels make no impact on the problem

The only way out for me when BFQ doesnt cut it is to rsync the data and use rsync's bwlimit parameter to "ease up" disk io a bit.

[Adarion]

Since we're mixing cpu govs and io scheds anyway, yes it still sucks. High disk I/O is a pain. Sometimes, on my small CF card powered systems I can't even log in to another console when on one is a high disc / flash card activity. You can type in your name + return; but the next prompt for the password is delayed, or the login process itself if you typed the password.
Though I rarely had problems in a GUI with the mouse, "just" CPU load (e.g. compiling on all cores) doesn't give me that much of a slowdown.

[liam]

First, this is something that has been a long time coming. I first heard about it in the context of big.LITTLE where it was made clear the disconnect between scheduler and governor led to very suboptimal results. Additionally, their separation made it very hard to create a sane implementation of CPU hotplug. So, hurray to Rafael!
Second, my understanding of Linux's WELL known issues with disk io and interactivity is that it is a multi headed beastie. One aspect is the elevator. Another is the CPU scheduler. A third is that Linux doesn't have a true async io implementation (and we probably won't have a one until nvm shows up, and even then they may go a different way). Forth, the dirty ratio should probably be dynamic. Fifth, we don't have a process swapping method, only a paging one.
This is why no (Linux)desktop is immune to this issue in all circumstances.

[c117152]

Turn off NCQ. And if you're already using BFQ, run Liquorix.