3 ms·

Is this a joke? Systems programming is a lot about accessing APIs, dealing with all sorts of intricacies like interrupts, different execution contexts, and mana

by gnull 4y ago

Is this a joke? Systems programming is a lot about accessing APIs, dealing with all sorts of intricacies like interrupts, different execution contexts, and managing memory as you said. If you write in C and your program is complex enough, you will spend a lot of time just chasing segfaults and concurrency bugs and getting it to work the first time you write it. If your systems programming is in userspace, that's sort of fine. But if you're in kernel or on bare metal, the cost of debugging goes up by an order of magnitude. There's no debuggers on bare metal, and nobody can tell you how your program crashed—the device just stops responding, that's it. That's why safe languages make even more sense in these restricted environments. If you get twice fewer memory safety bugs while you're getting your program to work, that can reduce your development time like 5-6 times.

jstimpfle 4y ago
> Systems programming is a lot about accessing APIs, dealing with all sorts of intricacies like interrupts, different execution contexts So now you need to make your interrupts talk to your Java objects? Is this any safer? Is it easier to get a VM running in your kernel (probably no mean feat to do that in the first place) and you'll never get any concurrency bugs? And if you reduce memory bugs by half, those will be easier to debug? And you won't be annoyed that you can't guarantee to be able to link objects in queues (because of allocation failure) and access them with generic code to copy data, link/unlink them, and so on? You're fine to pay for callbacks and interfaces everywhere, both in terms of runtime and performance as as maintenance headaches? I'm asking incredulously, but seriously. Because frankly I've never looked at a project like MirageOS or whatever. But given real world evidence of what has survived, I don't see why you should assume I'm joking. That you can't debug a "bare metal" kernel isn't quite true, either. But sure, the more complex a system becomes, the more contemplation it requires to figure out problems. This is universally true, but you can't simply discuss complexity away. And adding complex object models on top without consideration doesn't make your task easier just like that.
- gnull 4y ago
  I have a really hard time following anything you say here. How did we start talking about Java and VMs? Is this some sort of strawman argument? > But sure, the more complex a system becomes, the more contemplation it requires to figure out problems Strawman again? I wasn't talking about complexity. I said that the more "systems" your programming is, the higher is the cost of memory safety bugs.
  - jstimpfle 4y ago
    > How did we start talking about Java and VMs? Is this some sort of strawman argument? Is it a strawman if my comment was in response to "managed languages"? > Strawman again? I wasn't talking about complexity. I said that the more "systems" your programming is, the higher is the cost of memory safety bugs. For clarity, you spoke about the cost of debugging of memory bugs. And I said, it's universally true that programs are harder to debug the more "systems" they get. The reason is that it typically isn't sufficient to simply trace an individual thread anymore. "Logical tasks" are served over a number of event handlers executed in various (OS) threads. It's not first and foremost a refutal of what you said. But an observation that I even placed in opposition to my other statement that it's not quite true that you can't use debuggers with kernels. FWIW and so on. I don't get why you are calling "strawman" repeatedly, and don't get the aggressive tone of your comments.
- kaba0 4y ago
  An OS will always need some tiny assembly part, as some instructions needed for the kernel simply never gets generated by compilers. Also, an OS itself is pretty much a garbage collector for resources, it could very well reuse/link into its own GC for better performance. Are we really talking about the “price” of managed code, when C code is chock-full of linked list data structures? A list of boxed objects is more cache-friendly than that. It is simply not always that performance sensitive to begin with (e.g. does it really matter if it queries the available display connectors in 0.0001s or 10x that?) Regarding MirageOS, they actually achieved better performance by using a managed language than contemporary C OSs for select tasks. This is possible due to context switches being very expensive, and a managed env can get away with (some?) of those.
  - jstimpfle 4y ago
    Those linked lists give some nice guarantee that significantly simplify a lot of code and give some guarantees that you just can't have otherwise. Context switching can be expensive, but I don't see what's specific about managed envs about that. Fundamentally you have to have trust in code, and need to have the hardware support to enforce authenticity of the trusted code in order to avoid context switching. A different, promising development to reduce context switches is CPUs growing more and more cores, and more and more kernel resources being available through io_uring and similar async interfaces.

by gnull 4y ago

jstimpfle 4y ago

> Systems programming is a lot about accessing APIs, dealing with all sorts of intricacies like interrupts, different execution contexts So now you need to make your interrupts talk to your Java objects? Is this any safer? Is it easier to get a VM running in your kernel (probably no mean feat to do that in the first place) and you'll never get any concurrency bugs? And if you reduce memory bugs by half, those will be easier to debug? And you won't be annoyed that you can't guarantee to be able to link objects in queues (because of allocation failure) and access them with generic code to copy data, link/unlink them, and so on? You're fine to pay for callbacks and interfaces everywhere, both in terms of runtime and performance as as maintenance headaches? I'm asking incredulously, but seriously. Because frankly I've never looked at a project like MirageOS or whatever. But given real world evidence of what has survived, I don't see why you should assume I'm joking. That you can't debug a "bare metal" kernel isn't quite true, either. But sure, the more complex a system becomes, the more contemplation it requires to figure out problems. This is universally true, but you can't simply discuss complexity away. And adding complex object models on top without consideration doesn't make your task easier just like that.

gnull 4y ago
I have a really hard time following anything you say here. How did we start talking about Java and VMs? Is this some sort of strawman argument? > But sure, the more complex a system becomes, the more contemplation it requires to figure out problems Strawman again? I wasn't talking about complexity. I said that the more "systems" your programming is, the higher is the cost of memory safety bugs.
- jstimpfle 4y ago
  > How did we start talking about Java and VMs? Is this some sort of strawman argument? Is it a strawman if my comment was in response to "managed languages"? > Strawman again? I wasn't talking about complexity. I said that the more "systems" your programming is, the higher is the cost of memory safety bugs. For clarity, you spoke about the cost of debugging of memory bugs. And I said, it's universally true that programs are harder to debug the more "systems" they get. The reason is that it typically isn't sufficient to simply trace an individual thread anymore. "Logical tasks" are served over a number of event handlers executed in various (OS) threads. It's not first and foremost a refutal of what you said. But an observation that I even placed in opposition to my other statement that it's not quite true that you can't use debuggers with kernels. FWIW and so on. I don't get why you are calling "strawman" repeatedly, and don't get the aggressive tone of your comments.
kaba0 4y ago
An OS will always need some tiny assembly part, as some instructions needed for the kernel simply never gets generated by compilers. Also, an OS itself is pretty much a garbage collector for resources, it could very well reuse/link into its own GC for better performance. Are we really talking about the “price” of managed code, when C code is chock-full of linked list data structures? A list of boxed objects is more cache-friendly than that. It is simply not always that performance sensitive to begin with (e.g. does it really matter if it queries the available display connectors in 0.0001s or 10x that?) Regarding MirageOS, they actually achieved better performance by using a managed language than contemporary C OSs for select tasks. This is possible due to context switches being very expensive, and a managed env can get away with (some?) of those.
- jstimpfle 4y ago
  Those linked lists give some nice guarantee that significantly simplify a lot of code and give some guarantees that you just can't have otherwise. Context switching can be expensive, but I don't see what's specific about managed envs about that. Fundamentally you have to have trust in code, and need to have the hardware support to enforce authenticity of the trusted code in order to avoid context switching. A different, promising development to reduce context switches is CPUs growing more and more cores, and more and more kernel resources being available through io_uring and similar async interfaces.