Hacker News Reader

Comment by the_grue | Hacker News Reader

So WHY is Windows slower? What effect can OS conceivably have on CPU benchmarks? Off the top of my head I can think of preemptive multitasking and memory allocation implementations (the latter shouldn't influence pure CPU benchmarks much, but it can influence 7-Zip results), any better guesses?

View on HN

There was a very good post about that a few years ago. It probably doesn't explain this specific case, but it is a fascinating read nonetheless.

>I Contribute to the Windows Kernel. We Are Slower Than Other Operating Systems. Here Is Why

http://blog.zorinaq.com/i-contribute-to-the-windows-kernel-w...

Theres also the burden on Microsoft that nothing can break. Meaning: anything that used to work before (in an earlier incarnation of Windows) must continue to work in later versions. I don't think Linux has this burden. They are more free to break stuff and so more free to innovate.

I'm quite sure Linux has the same burden. Here is what Linus Torvalds has to say: https://lkml.org/lkml/2012/12/23/75

In Linux Kernel development there is one realy important Rule: Don't break Userspace!

Its like the Windows burden you mentioned, but much more reasonable imho.

Thanks, that's really interesting, even though the author toned much of his initial satire down in a later update.

curiously, nowadays (i.e. the article/original comment is from 2013) some google employees voice similar criticism about their own engineering culture: building new things are rewarded, maintaining an old thing is not.

I guess it's an industry-wide problem. It would be illuminating to learn of any counter examples, where a big software company manages NOT to fall into this trap.

Just a hunch, but maybe you could try searching outside of SV.

This deserves its own post.

it had one https://news.ycombinator.com/item?id=5688151

Scheduling threads is probably a big one I imagine. If Windows schedules two threads from the same process working on a shared dataset across two different NUMA nodes then memory latency can shoot up, from what I understand.

I've also noticed that Windows is very fond of passing hard-working threads around all the cores in turn, rather than letting them hog certain cores. Pretty sure that doesn't improve cache behavior etc.

You're running dozens of threads over 16+ cores who don't all have the same properties to access and share memory (speed, invalidation, permission, ...).

The OS is in charge of deciding who runs where, and who shares a core with who.

So ... A lot.

I wonder whether Windows Server SKUs do differently, perhaps due to different scheduler optimization choices.

There is no difference, in fact they might be slower as they are on older kernels.

Combination of things (which probably also explains why some distros are good at one becnhmark but less good at others)? Scheduler, no architecture-specific kernel code etc.

Plus it's never just the OS: there's always platform-specific source code and the compiler.. Non-optimal optimization options for the latter and you're screwed as well.

It might be blindly applying all the Intel brokeness (spectre/meltdown) patches to AMD cpu.

Reply on news.ycombinator.com