[akpa1]

Very cool idea, but for me it completely doesn't work. I'll open an image, the entire computer (running Linux) will briefly freeze (including stopping Spotify in its tracks) and when it comes back, the only thing that's left is a message from Firefox telling the tab has crashed.

[lxgr]

A website being able to do this is arguably a browser bug.

If a browser's sandbox can't even protect against accidental resource exhaustion, I'd be very concerned about that as an intentional attack vector.

[ipaddr]

What browser protects against resource exhaustion? Chrome and all variants do not. Running out of memory or cpu or even hard drive space can happen and does in all browsers.

[lxgr]

Every browser I know meters all APIs capable of using disk space, and 100% CPU usage doesn’t hang your system on any reasonable OS.

Memory can indeed be a problem, but at least if a tab becomes the largest single memory user on my system, the OOM killer will come for it first.

So if for CPU and memory browsers can lean on the OS for proper resource management but it’s not the case for GPU, maybe their WebGPU implementations aren’t ready for production yet.

[codedokode]

Before OOM killer activates, the system will swap out all other applications including desktop shell. If you are unlucky to use a spinning hard drive, your system will freeze for a long time.

This is a problem with popular Linux distributions, they do not have protection against application swapping out important system applications.

Linux has disk quotas and CPU scheduler but it doesn't have fair memory and swap management (or this is not configured out of the box). For example, desktop shell doesn't have protection against being swapped out.

[lxgr]

Ah yes, that is indeed very annoying. That's why I usually don't configure a swap partition on Linux, although that probably ends up wasting a lot of memory on write-once-read-never pages that could otherwise be used for the page cache.

A per-process real memory maximum that automatically invokes the OOM killer unless somehow opted out would probably be useful.

[codedokode]

Browsers do not protect from tab using all available memory and swapping out other applications including desktop shell.

[cmgriffing]

Looks like WEBGPU is only on nightly for FF, for now.

[jml7c5]

Sounds a bit like running out of memory, followed by the OOM killer going after the process... though I'm not sure the OOM killer would act that quickly. Do you get any interesting messages in `dmesg`?

[akpa1]

You're spot on - dmesg shows it's OOMing, and pretty fast too. Turns out that it happens even if I just open the web page, yet alone start trying to remove the background of an image, so I'd guess it's something to do with loading whatever model is being used.

[anduc]

I tried it on my Chrome-Mac device, but not sure how it works on other devices (assuming it works). I'll test it out later. Sorry for the inconvenience

[akpa1]

Please don't apologise for any inconvenience! It's a cool project and I'd love to see it working, I'm just sorry I can't give any positive feedback