I'm very surprised by the hack that reduces the area of possibles for a process to two CPUs. This will cause other problems when 32+ cores computers get more common.
I'm even more surprised by "some benchmarks show it's faster, let's merge it".
Maybe they could try something larger than subsets of 2 CPUs?
If I'm understanding correctly, the two core limit is only for initially waking the process, after which normal load balancing can move it elsewhere if necessary. There's a benefit to sticking to one of the two same hyper threaded CPUs in terms of cache, so this does make sense even on 32 core machines.
Yes, however I think you can get a 15-puzzle situation - even if you do this only for waking the process - when you wake a lot of process at the same time on a machine with many cores.
I really prefer Linus suggestion, even if it's the hardest.
Yeah, I really don't like blocking off all that extra capacity just because the OS is too lazy to find it. Keep a better data structure for fast searches, or improve the result over time instead of sticking with the first quick result. Reminds me of Android where background processes can never use the full resources of the device even if they are idle otherwise. They are in a scheduling class that simply doesn't allow it as a fixed percentage of total resources, not a smart one that gives them less priority. Now say an app comes around that wants to do some heavy media processing for the user in the background that the user requested be done. Oops, sorry, you don't get to use all your hardware. Same thing that is happening to the DBMS in this case.
In the Android example, I can see why they limit low-priority processes to a fraction of total resources: to avoid silently draining the battery. Android runs on phones, and using 100% CPU on a phone is a great way to kill battery life. Allowing higher usage would lead to users getting annoyed at their phones (and Android in general), when really blame would lie with poorly-written apps they installed.
But I agree with you about the Linux patch. Considering the performance impact, it seems irresponsible not to implement the best solution. The kernel isn't some web app where everyone gets the latest version automatically. Bad code in the kernel sticks around for a long time.
I'm even more surprised by "some benchmarks show it's faster, let's merge it".
Maybe they could try something larger than subsets of 2 CPUs?