Ask HN: Why can't OS have dedicated cores and memory?

detaro · on July 10, 2017

There are a lot of issues where dedicated cores don't help - e.g. locking issues (see the discussion today about windows and chromium builds)

twobyfour · on July 10, 2017

Would you mind explaining what some of these cases are and why it doesn't work that way, please? (Sorry for what might be a fairly elementary question.)

actuallyalys · on July 11, 2017

I'm not detaro, but I'll take a crack. It's kind of like traffic—if you're constantly stopping for other cars, it doesn't really help to have a faster car. If the UI thread is waiting for something else to happen, having an entire core dedicated to the UI just means that entire core is waiting. In the case detaro was alluding to, compiling Chrome was creating so many processes, the UI had to wait for the kernel to shut down processes before it could do what the UI needed it to do. [0]

Adit has some good posts on concurrency:

—http://adit.io/posts/2013-05-11-The-Dining-Philosophers-Prob...

—http://adit.io/posts/2013-05-15-Locks,-Actors,-And-STM-In-Pi....

[0]: https://randomascii.wordpress.com/2017/07/09/24-core-cpu-and...

davelnewton · on July 10, 2017

You can, at least in Linux. But in general the OS will do a better scheduling job than you will.

kpil · on July 10, 2017

Maybe more efficient, but the default settings put my actions in the same queue as everything else.

This can be rather annoying, especially if the machine is low on memory or swapping to death.

As a matter of design principle for a desktop machine, it should prioritize everything that I do.

That includes not stealing the keyboard focus when I'm typing, up to the point of rolling back any popups that opened just before I started to type.

jldugger · on July 10, 2017

Well, lets just say that there's a variety of levels a desktop environment, and not all of them are immediately obvious to the OS or hardware layers.

The linux kernel is designed to be generic enough to work on wifi routers, phones, laptops, and massive compute clusters. There's some thing you can do, but there's going to be a class of optimizations any general purpose OS must give up on. For example, Windows NT runs a lot of window manager stuff in kernel space (https://en.wikipedia.org/wiki/Architecture_of_Windows_NT). Once you pull that in, the scheduler will have a richer dataset work with. But it kinda makes 32MB routers and headless servers less viable. Even Server Core still has a window manager!

davelnewton · on July 10, 2017

> not stealing the keyboard focus when I'm typing, up to the point of rolling back any popups that opened just before I started to type

That's a wholly separate issue. On Linux, at least, this is configurable to an extent, although I haven't used it recently enough to be helpful. Dedicated cores are irrelevant, though.

avereveard · on July 10, 2017

Mostly because the os needs to coordinate everything and not all operations can be interrupted. Operating system maintain each program illusion to be alone on the hardware, virtualizing disk transfers, memory mappings, display regions and all the other hardware. Some operations once queued and dispatched need to complete before other operations can be accepted - that might happen for example when the os is figthing to access a memory page from disk but the disk is busy in a large dma transfer - then the illusion of being alone comes crashing down and no clever scheduling can fix it, because a class of operations without a guaranteed time constraint exist on the ibm pc platform.

Without a hard real time guarantee from the hardware, everything built on it cannot provide a guarantee on dispatching only the exact amount of operations that the system can handle.

And that happens even before accounting that the os or some driver within may have some cpu locking going on.

Bender · on July 11, 2017

This isn't specific to your question, but I have VM's that have 6 logical processors. I use taskset to set the affinity of all OS processes to logical proccesor 0. This doesn't work for some kernel threads, but it works for most processes. I then give haproxy a number of logical processors, or balance based on what is running on that VM.

On physical hardware I would use NUMA regions per set of applications as well to optimize the use of CPU cache and get lower latency to memory, but the overall memory availability will be reduced. This is done on some latency sensitive applications.

You can do similar things in Windows. There is probably a way in Mac, but I have never tried.

As others pointed out, there are many cases where doing all of this won't help, so it really depends on the problem you are trying to solve, or the optimization you wish to accomplish. There are certainly no one-size-fits-all solutions.

rb808 · on July 12, 2017

https://en.wikipedia.org/wiki/Processor_affinity

In my experience we used affinity with older CPUs to improve single thread performance. Recently it doesn't make much difference.

uptownhr · on July 10, 2017

Feels like there isn't a solution using current OSes.

Why does CPU locking exist? I'm sure this is pretty technical question but if anyone has a dumb answer for me... Is this something impossible get rid of? Can it be scoped to a core and not the entire CPU?

zamalek · on July 10, 2017

iOS supposedly runs the UI thread at maximum priority on phones, which is similar to what you are suggesting here.