Hacker News Reader

Comment by highwaylights | Hacker News Reader

The problem with each of these tools (Lima etc.) is that it’s still fundamentally a virtual machine under the covers.

The great thing for me about WSL (and why IMHO it was worth all the effort they put into it), is that because the kernel process tree is running as a native Windows process tree, I don’t have to pre-allocate memory ahead of time.

This is murder on Apple’s RAM-restricted laptops, and kind of rubbish on machines with massive RAM too. Most of the time, most of your memory goes unused. The base model Airs/Pros for example need to have 25-50% of your RAM locked away just to run a Linux container at all - when you already have a fairly anaemic 8GB to begin with that’s pretty dire.

View on HN

It's a shame they essentially deprecated WSL1 and is doing everything in WSL2 now, where exactly this kind of memory allocation and the performance issues that entails is an issue.

Not that I think it's a bad decision, but I wish they somehow managed to break Conway's Law and enhance Windows itself by fixing the file system speed issues as well as somehow making NT/Windows processes fork instantaneously just like on Linux :)

Their explanation was that Windows could not be enhanced but rather should be slimmed down - the APIs that the NT kernel provides can't be efficiently worked around without breaking all kinds of already existing Windows software.

We work very closely across the file system, kernel, and virtualization teams, so I don’t think we can blame Conway’s Law for WSL2. And actually our WSL1 fork performance wasn’t too bad IIRC—we have real fork at the Windows kernel level, it’s just not something that can realistically work with the Win32 programming model. I also think we will eventually resolve the file system performance issues.

No, the most important reason to choose a VM instead of a reimplementation of the Linux ABI is long tail compatibility. You can’t realistically replicate and then keep up with every corner of the Linux kernel’s interface. And so with WSL1, software will randomly not work, or it will randomly break after an apt upgrade, and users will get frustrated and switch to a VM anyway. Might as well get perfect compatibility and still have nice integration with Windows via the WSL2 approach.

Thanks for the detailed explanation! You folks have actually tried to implement an ABI compatible Linux kernel, so I'll definitely take your word for it :)

I think most people understand why the WSL2 approach eventually won out, and its saved my bacon more times than I can count when I had to demo something and it breaks on my Mac only for me to quickly switch to "real" linux on my personal windows machine running WSL2.

However, with WSL1, implementing the Linux ABI was such a wicked flex.

Seconding that. WSL1 is an impressive achievement. I use WSL every day and switched to WSL2 a long while back but I still kinda wish WSL1 were the way forward.

> when I had to demo something and it breaks on my Mac only for me to quickly switch to "real" linux on my personal windows machine running WSL2.

How's that any different than running a virtual machine on your Mac with a real Linux in it?

Seamless mounting, launching code editors in Windows from Linux, issuing docker commands in the Linux OS while the docker engine and its mounts are hosted on Windows.

Doing things with something like Vagrant on Mac sometimes cuts it, sometimes doesn't. Either way, if I'm not shelling out money I could be stuck using a subpar solution like VirtualBox.

You can’t realistically replicate and then keep up with every corner of the Linux kernel’s interface

Didn't Linus famously say they'll "never break userspace"?

New syscalls and flags get added. The Linux ABI is huge though, even if you snapshot at some “old” 4.x version, say. WINE has similar problems with coverage, and Win32 is famously “stable” too.

Yes. But user space can take dependencies on features we didn’t implement yet or we implemented incorrectly or that have bugs in combinations with other features. E.g., more software is starting to depend on namespace features that we implemented incompletely in earlier versions of WSL1. So Docker worked for a while until they starting using a different subset of kernel features, for example.

At another point, glibc started depending on more precise behavior of CLONE_VFORK, which we originally didn’t implement fully. So essentially all of user space was broken. We fixed it as soon as we could, but I think glibc may have added a workaround, too. I feel bad that the community had to work around our bugs.

WSL2 is better. I'll take correctness over a slight loss in performance.

Also Linux is very fast. So it's not like there is a huge performance loss in most cases, I am guessing.

Doesnt ReFS do exactly that? I have been surprised they do not try to push it more - the fact it cannot (or maybe it can now but in the past it could not) be used for a boot drive seemed strange.

> No, the most important reason to choose a VM instead of a reimplementation of the Linux ABI is long tail compatibility. You can’t realistically replicate and then keep up with every corner of the Linux kernel’s interface.

As far as we the public know, Google Cloud Run is based on gVisor, which emulates Linux system calls with userspace code. Seems to work great for the usual container workloads.

The other commenter alluded to this but didn’t say it outright: WSL2 is essentially just a special VM with API based limited hardware access. It’s neat though, the file systems are connected over 9p.

It is - but its speciality is exactly what matters most - it doesn't preallocate RAM and in fact can dynamically release it back to the host without any user intervention. It's super convenient.

To this day I wish they'd release decent (i.e. well documented) APIs for creating VMs like this yourself with Hyper-V. I sometimes have to spin a bunch of Windows VMs for testing, and if they could dynamically release RAM and CPU when they aren't using it it would be amazing. They have something like it with Windows Sandbox but you can only spin one at a time.

There are PowerShell functions for creating VM’s.

That's not what OP meant. OP meant interfacing with Hyper-V from inside the VM.

Things like dynamic memory allocation and release aren't some kind of black magic, but are something that is actually directly communicated by the guest VM.

Hyper-V does support memory ballooning (dynamic memory allocation from host to VM) - is that what you mean, or do I misunderstand?

This does need support in the guest VM, but I've only ever had issues with really old Linux distros.

My comment was really kind of inaccurate, the difference isn't the memory ballooning, it's the special way WSL2 is able to interact with the host. It frees memory much better than normal Hyper-V VMs, and boots way faster than a regular Ubuntu VM.

Microsoft implemented a Firecracker style microVM (they call it "Krypton") for Hyper-V, which is used for WSL2, Edge Application Guard (run Edge in a VM), and Windows Sandbox. There aren't any real good docs on how to create one yourself using the Host Compute Service.

That makes sense, thanks for explaining!

To add to this a little bit. There is deeper host side integration for these applications and the VM’s are virtual address backed (the guest vm’s memory is represented as a process on the host), which adds more opportunities to free up physical memory. The windows sandbox architecture page covers some of this.

https://learn.microsoft.com/en-us/windows/security/threat-pr...

Also WSL2, if I recall correctly, has a relatively minimal base OS with a custom init binary.

I would say if you’re looking to use Host Compute Service, your best option would be to look at the way https://github.com/microsoft/hcsshim initializes containers.

That's been a feature of existing VM software for years.

CoLinux did that years before WSL1 did. Sadly, the project is dead and AFAIK it never really worked on Windows 64-bit, only on 32-bit: https://en.wikipedia.org/wiki/Cooperative_Linux

You are right and it can definitely be improved, however at all my jobs I’ve had 32-64gb MacBooks, is that not true for most developers? When I didn’t i was able to buy an actual Linux laptop

Depends on the company and developer I suppose.

Web dev here, ran with 16 up until recently… now at a whopping 24…

No memory pressure.

Nope, most contractors get issued a laptop with 8GB unless their project is doing anything container related, in which case they get a 16GB one.

Assuming that they just don't get a cloud VM accessed via Citrix/RDP.

Reply on news.ycombinator.com