6 ms·

>why in the heck is he going for POSIX compatibility Because existing desktop applications can be ported to ToaruOS >why not a safer microkernel, keeping ever

>why in the heck is he going for POSIX compatibility Because existing desktop applications can be ported to ToaruOS >why not a safer microkernel, keeping everything in userspace? This is a design choice, microkernels aren't necessarily better than hybrid, they're slower, harder to debug and process management can be complicated

fouc 5y ago
Isn't QNX a microkernel? I remember it being known for being quite fast?
- guenthert 5y ago
  As a real-time OS it is known for deterministic response times. If it were exceptionally fast (and licenses cheap enough), you'd see hosts in the TOP500 using it.
  - eggy 5y ago
    I agree 100%. QNX is lurking in products you may use. It was the OS for the show control system that is used throughout entertainment where real-time is necessary for the safety of the devices it controls, which also have hardware safety at the lower level. I would drop into the QNX terminal for certain tasks. Unfortunately, you used to be able to download the show control software and play with it, but it has since been bought buy a company that sells it with the equipment they rent, so you need to buy trainging and it is behind their wall now. Not QNX, but the show control software that runs on QNX.
  - p_l 5y ago
    TOP500 is chock full of microkernels though, even if the "I/O nodes" would run Linux
    - guenthert 5y ago
      I'd like to see more information about that. I remember that Penguin Computing offered such some 15 years ago, but don't know where it was deployed or still is. Cray and IBM had also such a concept for their superclusers in the past, but are they still using such? The one HPC environment I worked on (a major car manufacturer in Europe) used plain RH Linux on all nodes as recently as three years ago. The current #1 (Fugaku) uses IHK/McKernel as kernel for the actual payload. The previous #1 (IBM Summit) seems to use RH Linux though. Perhaps, since the most performance critical part is run by and within the GPGPU(s), the actual OS doesn't matter all that much (for performance -- it matters of course for programmer's comfort/efficiency).
      - p_l 5y ago
        There used to be a lot of "special microkernel on compute RPCing to Linux on I/O" on Crays and the like. Hard to say how prevalent it is now, and most annoyingly I can't recall the names. (Charon?)
- bregma 5y ago
  No, it's more of a nanokernel. It's very fast. Full disclosure: I maintain QNX toolchain.
md8z 5y ago
I'm very confused by this comment. There are a ton of other things you need to implement if you want to have desktop applications. POSIX does not specify any APIs for graphical applications. You might be thinking of something else. If you want to support the lion's share of desktop applications, it would actually be better to implement the Win32 API...
- craigmart 5y ago
  Sorry, I meant software in general but wrote "desktop applications" instead. Anyway the sentence is still valid, even if you'll have to implement other things such as the graphical interface, the POSIX compliant code won't need modification
  - md8z 5y ago
    If you're taking an app built for Linux or GNU or BSD, then it probably will need modification, as those systems have various extensions on top of POSIX.
eggy 5y ago
Just curious how hard it would be to forego POSIX entirely if you were building an OS. I know TempleOS is entirely from scratch. I'd like to implement a small LISP like SectorLISP [1] (see yesterday's posts too on HN). I don't know much about building my own OS, so I'd like to start with something like MenuetOS (my first PL was asm), SerenityOS, TempleOS, or this one. I'd like it to be completely an 'island', i.e. POSIX not a requirement. I want to use it to hack on in isolation without any easy copy/paste shortcuts. I know Mezzano exists, and it has booted on bare metal, but I would like to start with the OS's above, implement my own LISP, and go from there. Any other OS recommendations base on my ignorant, but wishful, reqs above? I realize there are some others in Rust too. Thanks! [1] https://github.com/jart/sectorlisp https://github.com/jart/sectorlisp
- adrian_b 5y ago
  Someone who would make a new OS, should define a completely new system call interface, as it is likely that now it is possible to conceive a better interface than 50 years ago and anyway if it would not be different there would be no reason to make a new OS, instead of modifying an existing one. Nevertheless, the first thing after defining a new OS interface must be writing a POSIX API translation layer, to be able to use without modifications the huge number of already existing programs. Writing a new OS is enough work, nobody would have time to also write file systems, compilers, a shell, a text editor, an Internet browser and so on. After having a usable environment, one can write whatever new program is desired, which would use the new native OS interface, but it would not be possible to replace everything at the same time. Besides having a POSIX translation layer, which can be written using as a starting point one of the standard C libraries, where the system calls must be replaced with the translation layer, some method must be found for reusing device drivers made for other operating systems, e.g. either for Linux or for one of the *BSD systems. Nobody would have time to also write all the needed device drivers. So there must exist some translation layer also for device drivers, maybe by running them in a virtual machine. The same as for user applications, if there is special interest in a certain device driver, it should be rewritten for the new OS, but rewriting all the device drivers that could be needed would take years, so it is important to implement a way to reuse the existing device drivers.
yjftsjthsd-h 5y ago
> microkernels aren't necessarily better than hybrid, they're slower, harder to debug and process management can be complicated I was basically on board, but how are they harder to debug? I'd think being able to run components in userspace would make debugging way easier.
- mjgerm 5y ago
  You are now debugging a distributed system.
  - jatone 5y ago
    fun fact: you already are in linux. being a monolith doesn't change the nature of the problem.
    - yjftsjthsd-h 5y ago
      Are you? I'm paddy pretty sure I can run a single Linux and point a single gdb at it[0] and debug it in a single memory space; I don't think you can do that with a microkernel. [0] possibly resorting to UML, but still
  - yjftsjthsd-h 5y ago
    Oh, good point; I was thinking at the component level

fouc 5y ago

Isn't QNX a microkernel? I remember it being known for being quite fast?

guenthert 5y ago
As a real-time OS it is known for deterministic response times. If it were exceptionally fast (and licenses cheap enough), you'd see hosts in the TOP500 using it.
- eggy 5y ago
  I agree 100%. QNX is lurking in products you may use. It was the OS for the show control system that is used throughout entertainment where real-time is necessary for the safety of the devices it controls, which also have hardware safety at the lower level. I would drop into the QNX terminal for certain tasks. Unfortunately, you used to be able to download the show control software and play with it, but it has since been bought buy a company that sells it with the equipment they rent, so you need to buy trainging and it is behind their wall now. Not QNX, but the show control software that runs on QNX.
- p_l 5y ago
  TOP500 is chock full of microkernels though, even if the "I/O nodes" would run Linux
  - guenthert 5y ago
    I'd like to see more information about that. I remember that Penguin Computing offered such some 15 years ago, but don't know where it was deployed or still is. Cray and IBM had also such a concept for their superclusers in the past, but are they still using such? The one HPC environment I worked on (a major car manufacturer in Europe) used plain RH Linux on all nodes as recently as three years ago. The current #1 (Fugaku) uses IHK/McKernel as kernel for the actual payload. The previous #1 (IBM Summit) seems to use RH Linux though. Perhaps, since the most performance critical part is run by and within the GPGPU(s), the actual OS doesn't matter all that much (for performance -- it matters of course for programmer's comfort/efficiency).
    - p_l 5y ago
      There used to be a lot of "special microkernel on compute RPCing to Linux on I/O" on Crays and the like. Hard to say how prevalent it is now, and most annoyingly I can't recall the names. (Charon?)
bregma 5y ago
No, it's more of a nanokernel. It's very fast. Full disclosure: I maintain QNX toolchain.

md8z 5y ago

I'm very confused by this comment. There are a ton of other things you need to implement if you want to have desktop applications. POSIX does not specify any APIs for graphical applications. You might be thinking of something else. If you want to support the lion's share of desktop applications, it would actually be better to implement the Win32 API...

craigmart 5y ago
Sorry, I meant software in general but wrote "desktop applications" instead. Anyway the sentence is still valid, even if you'll have to implement other things such as the graphical interface, the POSIX compliant code won't need modification
- md8z 5y ago
  If you're taking an app built for Linux or GNU or BSD, then it probably will need modification, as those systems have various extensions on top of POSIX.

eggy 5y ago

Just curious how hard it would be to forego POSIX entirely if you were building an OS. I know TempleOS is entirely from scratch. I'd like to implement a small LISP like SectorLISP [1] (see yesterday's posts too on HN). I don't know much about building my own OS, so I'd like to start with something like MenuetOS (my first PL was asm), SerenityOS, TempleOS, or this one. I'd like it to be completely an 'island', i.e. POSIX not a requirement. I want to use it to hack on in isolation without any easy copy/paste shortcuts. I know Mezzano exists, and it has booted on bare metal, but I would like to start with the OS's above, implement my own LISP, and go from there. Any other OS recommendations base on my ignorant, but wishful, reqs above? I realize there are some others in Rust too. Thanks! [1] https://github.com/jart/sectorlisp https://github.com/jart/sectorlisp

adrian_b 5y ago
Someone who would make a new OS, should define a completely new system call interface, as it is likely that now it is possible to conceive a better interface than 50 years ago and anyway if it would not be different there would be no reason to make a new OS, instead of modifying an existing one. Nevertheless, the first thing after defining a new OS interface must be writing a POSIX API translation layer, to be able to use without modifications the huge number of already existing programs. Writing a new OS is enough work, nobody would have time to also write file systems, compilers, a shell, a text editor, an Internet browser and so on. After having a usable environment, one can write whatever new program is desired, which would use the new native OS interface, but it would not be possible to replace everything at the same time. Besides having a POSIX translation layer, which can be written using as a starting point one of the standard C libraries, where the system calls must be replaced with the translation layer, some method must be found for reusing device drivers made for other operating systems, e.g. either for Linux or for one of the *BSD systems. Nobody would have time to also write all the needed device drivers. So there must exist some translation layer also for device drivers, maybe by running them in a virtual machine. The same as for user applications, if there is special interest in a certain device driver, it should be rewritten for the new OS, but rewriting all the device drivers that could be needed would take years, so it is important to implement a way to reuse the existing device drivers.

yjftsjthsd-h 5y ago

> microkernels aren't necessarily better than hybrid, they're slower, harder to debug and process management can be complicated I was basically on board, but how are they harder to debug? I'd think being able to run components in userspace would make debugging way easier.

mjgerm 5y ago
You are now debugging a distributed system.
- jatone 5y ago
  fun fact: you already are in linux. being a monolith doesn't change the nature of the problem.
  - yjftsjthsd-h 5y ago
    Are you? I'm paddy pretty sure I can run a single Linux and point a single gdb at it[0] and debug it in a single memory space; I don't think you can do that with a microkernel. [0] possibly resorting to UML, but still
- yjftsjthsd-h 5y ago
  Oh, good point; I was thinking at the component level