The whole concept of "virtual memory" has tainted systems design for decades. Treating RAM as a cache relies on the OS making guesses about what will be needed and what can be passivated without it actually knowing the application requirements. Except that compared to CPU level caching, the cost of page faults is big enough that performance degradation is not linear and breaks the user experience. The idea that a 4GB machine can do the same with as an 8GB one albeit slower is just not true. If you hit the swap, you feel it bad. I'll concede that Zram can work because the degradation is softer. But anything hitting the IO should be explicitly controlled by the app.
Other random semi-related thoughts:
- Rust having to define a new stdlib to be used in Linux kernel because of explicit allocation failure requirements. Why wasn't this possibility factored in from the beginning?
- Most software nowadays just abstracts memory costs, partly explaining why a word processor that used to work fine with 64mb of RAM now takes a gig to get anything done.
- Embedded development experience should be a requirement for any serious software engineer.
> Rust having to define a new stdlib to be used in Linux kernel because of explicit allocation failure requirements.
This is phrased in a way that’s a bit more extreme than in reality. Some new features are in the process of being added.
> Why wasn't this possibility factored in from the beginning?
So, there’s a few ways to talk about this. The first is… it was! Rust has three major layers to its standard library: core, alloc, and std. core, the lowest level, is a freestanding library. Alloc introduces memory allocations, and std introduces stuff that builds on top of OS functionality, like filesystems. What’s going on here is the kernel wanting to use the alloc layer in the kernel itself. So it’s naturally a bit higher level, and so needs some more work to fit in. Just normal software development stuff.
Why didn’t alloc have fallible APIs? Because of Linux, ironically. The usual setup there means you won’t ever observe an allocation failure. So there hasn’t been a lot of pressure to add those APIs, as they’re less useful then you might imagine at first. And it also goes the other way; a lot of embedded systems do not allocate dynamically at all, so for stuff smaller or lower level than Linux, there hasn’t been any pressure there either.
Also, I use the word “pressure” on purpose: like any open source project, work gets done when someone that needs a feature drives that feature forward. These things have been considered, for essentially forever, it’s just that finishing the work was never prioritized by anyone, because there’s an infinite amount of work to do and a finite number of people doing it. The Rust for Linux folks are now those people coming along and driving that upstream work. Which benefits all who come later.
Oh hello, thanks for the clarification! Having enjoyed writing some embedded Rust, I'm familiar with the core/alloc/std split. IIUC you're saying that the user-space Linux malloc API itself does not provide a reliable way for the application to think about hard memory limits? Which would fuel my pet theory about "infinite virtual memory" being a significant factor in the ever growing software bloat.
Ah, okay. So yeah, it's not a new standard library, it's "things like Vec are adding .push_within_capacity() that's like push except it returns a Result and errors instead of reallocating" more than "bespoke standard library."
> IIUC you're saying that the user-space Linux malloc API itself does not provide a reliable way for the application to think about hard memory limits?
So, unless you've specifically configured this to 2, there are many circumstances where you simply will not get an error from the kernel, even if you've requested more memory than available.
What happens in this case is that your program will continue to run. At some point, it will access the bad allocation. The kernel will notice that there's not actually enough memory, and the "oom killer" will decide to kill a process to make space. It might be your process! It also might not be. Just depends. But this happens later, and asynchronously from your program. You cannot handle this error from inside your program.
So even if these APIs existed, they wouldn't change the behavior: they would faithfully report what the kernel reported to them: that the allocation succeeded.
Most of the time, you want to use RAM as a cache for the disk. I was trying to make the argument that sometimes that disk cache is more valuable than an under-used anonymous mapping.
Steve has responded to your comment about Rust; to your other comments:
Modern applications do a lot more than old ones. Even if you only use 20% of the features, you probably use a different 20% from any arbitrary other person. You also probably benefit from the OS being able to map everything into virtual memory but only actually load the bits you use :).
And I strongly disagree with your stance on being "serious". I'm sure you don't mean to gate-keep, but we need to teach people where they are rather than giving them hoops to jump through.
In my experience, some of the best software engineers have very little development background. And I say that as someone who implemented 64-bit integer support for the compiler and RTL for a DSP part back in the day. It's useful to have people around with a variety of backgrounds, it's not necessary for everyone to share any particular experience.
Other random semi-related thoughts:
- Rust having to define a new stdlib to be used in Linux kernel because of explicit allocation failure requirements. Why wasn't this possibility factored in from the beginning?
- Most software nowadays just abstracts memory costs, partly explaining why a word processor that used to work fine with 64mb of RAM now takes a gig to get anything done.
- Embedded development experience should be a requirement for any serious software engineer.