CFS tweaks might hurt performance about system76-scheduler

[mmstick]

Want to help out with a PR?

[xalt7x]

I'm afraid it will take too much time for me to compile, test and make a PR.

--

Found a good source for the first suggestion (""minimum_granularity" value). It's right in the linux kernel source code (kernel/sched/fair.c):

/*
 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
 */
static unsigned int sched_nr_latency = 8;

[mmstick]

The default sched_latency is 6ns, and 6ns / 8 = 0.75, so the default that we have with a min granularity of 0.75ns is correct. But we can drop the migration cost change.

[xalt7x]

The default sched_latency is 6ns, and 6ns / 8 = 0.75, so the default that we have with a min granularity of 0.75ns is correct.

But in the current readme (and even source code) of system76-scheduler there is suggested "minimum_granularity" value for "Responsive" CPU Scheduler Latency Configuration which is not correct for the most distros.

latency: 4ns
minimum_granularity: 0.4ms

[mmstick]

That is because the responsive config is the Zen config, which is much more ideal for desktop responsiveness than what most distributions have as their defaults. There wouldn't be any point in having a responsive config if it had the same values the default config. latency 4 / nr_latency 10 = 0.4 minimum granularity.