Fedora Asahi Remix | Hacker News Reader

Fedora Asahi Remix

bentsukun.ch

22 pointswalterbell19 comments

> Using Fedora on this hardware feels super snappy. File system operations are shockingly fast. Compiling software seems much faster than on macOS, though I did not benchmark this. My hunch is that the process management is a bit slower in macOS, due to the Mach / XNU underpinnings. 3D graphics performance is good too.

I'm glad that people are starting to notice this, Linux has been great for file system operations for a long time now.

I have a theory about why that is, and I think it's because of Linus Torvalds' obsession with performance, this is why git is so fast as well.

Torvalds has said multiples times that he is a performance centric person and it shows:

https://www.youtube.com/watch?v=MShbP3OpASA&t=1335s

https://www.youtube.com/watch?v=MjIPv8a0hU8&t=3027s

I also believe that it's going to be impossible to beat Linux in the performance game as time goes on, because things like io_uring and similar optimizations are just going to make other systems pale in comparison.

Single-threaded (in the kernel/control flow sense, not posix threads) nature of filesystem server in Minix v2 is presented as major point of why Linux is monolithic - by having all of I/O including FS handling in the monolithic kernel, Linus got multithreaded filesystem "for free" (sans locking, which before SMP was less of an issue too)

Interesting, I didn't know that. Thanks for sharing!

>Linux has been great for file system operations for a long time now

Doesn't that performance vary wildly between the different Linux filesystems?

> Doesn't that performance vary wildly between the different Linux filesystems?

To some degree it does, I'd say it depends a lot in your use case.

Linux is still going to feel snappy because of other optimizations though, even with a slow disk and filesystem.

The way Linux manages RAM is also important as everything is cached by default.

I'd love to hear technical explanations of the differences in performance for the most popular Linux filesystems (ext4, XFS, BTRFS, ZFS) vs. APFS and ZFS on macOS. But differences are certainly very noticeable between the ext4 Linux VMs on my personal Mac and the native APFS filesystem for anything that touches lots of files (compiling big programs, basically any Nix commands).

I've not thought about performance specifically, but I've always thought of Linux's great filesystem support in terms of breadth, features, and activity of development as a consequence of Linux's place in the server world. People run storage at incredible scale on Linux, and then mere desktop users can take advantage of the same tech on their laptops and it works pretty well. I think that makes some sense for performance, too: people run databases where IO efficiency eventually translates to money saved every month, and then we desktop users benefit from many of those as well.

I imagine that if Apple were still targeting the server market and it were a big part of their revenue, filesystem features and performance on macOS would look more like how they do on Linux.

> I think that makes some sense for performance, too: people run databases where IO efficiency eventually translates to money saved every month, and then we desktop users benefit from many of those as well.

Indeed, that's a huge advantage of Linux, the fact that we all run the same Linux kernel and benefit from improvements coming from all directions.

Linus himself mentions this advantage here: https://www.youtube.com/watch?v=MShbP3OpASA&t=2330

There's a nice recent benchmark here (Bcachefs vs. Btrfs vs. EXT4 vs. F2FS vs. XFS): https://www.phoronix.com/review/linux-611-filesystems

Bcachefs is starting to look really good against Btrfs, no APFS or ZFS though, but still interesting nonetheless.

It's been a hot minute since I've seen asahi. How stable is it? Is it a solid alternative if you're already used to Linux? I don't mind MacOS at all, but this seems like a best-of-both-worlds

I've been dual-booting it on a MBP and it's great. I use it for embedded Rust development and haven't had a single problem - it could be used in a production environment, even with all the dongles you need since the MBP is USB-C only.

Two complaints:

1. System updates take a lot longer than it feels like they should be taking. And if you are using the GUI it looks like things have frozen. The CLI is better from a knowing-whats-going-on point of view, but it still can use improvements.

2. The notch. Say whatever you want about the notch, but you've got usable pixels on either side of it. I want to use them. Without them, the screen is 16:9 or worse (???).

You can enable the pixels around the notch via a boot parameter. Yes, it's way better.

Oh wow, I will have to look that up! Thank you!

I have this in my /etc/defaults/grub:

GRUB_CMDLINE_LINUX_DEFAULT="quiet apple_dcp.show_notch=1"

Progress had been fast the first year but from a purely end-user point of view it stalled last year. It still can't last a day on standby, no HDMI, no thunderbolt, and some more problems I care less about.

It's still a hobby toy.

I liked fedora but had to give it up because I couldn't figure out how to pick a mirror somewhat permanently. It insists on contacting mirrors around the continent at random which my firewall blocks, and I prefer. Many searches, never found the definitive answer to pick a mirror.

They have centralized the mirrors in something called MirrorManager - ironically, it makes selecting Just One more difficult.

The intention is for parameters like 'minrate' to be set in DNF so that it may select to change mirrors... if rates aren't sustained.

If, for example, the download speed falls lower than 10Mbit over thirty seconds, it can pivot. Both parts of the formula are configurable - time period/rate.

In True Stallman spirit, they want the systems to work on your behalf, not the other way around. Keep many mirrors so that the client can effectively get the best one.

Sometimes 'best' means the fastest or most capable to meet the requested rates, other times... simply available. Unless you're running the mirror you probably don't want to limit yourself.

To that end... one can point networks of clients at specific mirrors using netblocks, but this is more for campus administrators:

https://fedoraproject.org/wiki/Infrastructure/MirrorManager#...

If still committed to one mirror, there is nothing stopping the replacement of 'metalink' lines in the repo definitions with 'baseurl'. That takes out MirrorManager

Yes, had a feeling it had to do with the ini files in dnf.d, however there were several of them with several sections each. Started writing a sed script to change them but result didn’t work. Gave up and went back to mint since it finally updated. :-/

Anyone knows if pytorch on asahi linux can use the MPS backend?

Or maybe more generically, are the apple silicon graphics drivers enabling the Metal API on linux?

No, you don’t get Metal on Linux.

They recently got Vulkan working, so if something can target the subset of Vulkan compute that is supported, it might work, but it’s fairly recent and afaik very few ML frameworks support a Vulkan backend.

Reply on news.ycombinator.com