New Patches Aim To Optimize Context Switching With Two Improvements

[phoronix]

Phoronix: New Patches Aim To Optimize Context Switching With Two Improvements

A set of Friday night patches provide for some exciting context switching optimizations to the Linux kernel...

https://www.phoronix.com/news/Linux-2024-Optimize-Ctx-Switch

[Errinwright]

Good stuff

[coder]

This warning never ever seems to fire, even on a very large fleet, so it may be best to hide that behind CONFIG_DEBUG_VM. With the web server workload, this is also about 17% of switch_mm_irqs_off."

Heh, a couple years from now, someone will probably discover a performance-robbing scheduler bug that's context-switching lots of threads on the same core, and it will turn out that this warning could've caught it early. Especially with people experimenting with sched_ext, we will want such warnings.

Now, I'm not saying to leave it always enabled. However, I wouldn't hide it behind a config option that's almost never used or that someone wouldn't naturally know to use, while tinkering with schedulers.

[intelfx]

Heh, a couple years from now, someone will probably discover a performance-robbing scheduler bug that's context-switching lots of threads on the same core, and it will turn out that this warning could've caught it early. Especially with people experimenting with sched_ext, we will want such warnings.

I'm not sure how this warning (in essence an internal consistency assertion in the TLB management code) would aid in "catching early" any kind of a high-level scheduler logic bug?

[coder]

I'm not sure how this warning (in essence an internal consistency assertion in the TLB management code) would aid in "catching early" any kind of a high-level scheduler logic bug?

Isn't it essentially informing you that you're context-switching to/from the same core? Someone clearly thought it was a useful check to have, at one point in time. Maybe it's now redundant with a similar check at a higher level, but it seems like you'd want to have such a sanity-check to be sure the scheduler isn't generating a bunch of unnecessary/nop context switches.

[ayumu]

Still dogshit slow next to seL4's context switch.