His previous attempt on the same policy was https://github.com/hamadmarri/popcorn

This new one looks simpler and a bit better.

It would be interesting to see how this scales with cores/threads. Current gen CPUs have ~128 threads per socket and you can find hardware that is quad socket. I've seen some systems with <512 concurrent processes running. In those use cases would this patch have the effect of pinning each process to a single thread (ignoring io/other sleep states)? Would the performance benefits of this patch scale super-linearly against thread count?

"No support for NUMA, NUMA must be disabled." means you probably don't want to use this with a multi socket system. It looks like idle is enqueued when no tasks are active and (in the case where hardware threads > requested threads) for something where the task is always busy it'd stay on the current core (though that's not particularly unique to this scheduler).

Also worth noting "This scheduler is designed for desktop usage since it is about responsiveness. It may be not bad for servers.".

Very nice! I hope this is considered for merge into the kernel as an optional scheduler.