Hacker News Reader

Comment by dreamcompiler | Hacker News Reader

Even without tail recursion, conventional procedure calls are still just gotos plus an automatically-managed stack. Which fits the author's thesis that higher-level memory models are what distinguishes higher-level programming.

The aspect of this that bothers me is that automatically-managed stacks are not the only way to call procedures. Continuations for example are another way; actors yet another. But because automatically-managed stacks are present in virtually every CPU at the ISA level, many programmers never venture outside that model and have trouble understanding that others even exist, much less what advantages they might bring.

View on HN

If it makes you feel any better, I'd bet you all 3 UFOs still use call stacks. Even aliens are stuck on 8086 esque CPUs ;)

And what are those advantages? Function call ABIs are a standard way to change what is executing and come back when it is done. Does that really need to change? Other software architecture can be built on top of that, you don't need to change the way you call a function with assembly to structure software differently.

The advantages are that you can do threads, exceptions, lazy sequences, backtracking, tail-call elimination for state machines, and so on. (The disadvantages are the same.) Yes, you can emulate these on top of function-call-based ABIs, but you incur the usual emulation overhead.

this sub-discussion about tail calls, etc. is in reference to:

https://dspace.mit.edu/bitstream/handle/1721.1/5753/AIM-443....

that series of papers was released adjacent to the development of the scheme language which will generally optimize out the need for a call stack and replace it with essentially goto in certain cases.

as for 'changing the way you call a function' this subtopic is primarily of importance for language/compiler implementers; while you don't "need" to change the way you call a function, often doing so can speed up the program hugely, or allow things that were not otherwise possible (e.g. infinite recursion without stack overflow)

Tail call elimination isn't always possible, but it's the gateway drug to continuations which are always possible, and continuations are the gateway drug to actors, and once you have actors you're free to move any element of your program anywhere, break it into pieces and put them back together, go nuts, it just works--and now you're able to write code on asynchronous million-node networks without having to worry about locks, mutating global data, shared stuff, and everything else that gets in the way of using modern hardware.

If you have code that doesn't need to mutate data, it is already easy-ish to write on whatever scale you want. The problem is that there are very few algorithms and business problems that don't require shared state to work.

I'm not sure how actors or continuations could solve the problem of creating a distributed database of clients, or even of computing histograms on data in parallel, significantly more easily than more traditional programming paradigms.

Rather than accessing the shared data directly, you send a message to the actor that owns that piece of shared data, that message could be a request for some data, or a request to modify it in some way.

The actor that owns the data reads its mailbox of requests and processes them without worrying about locks and things, because it has exclusive access.

The actor that needs to access the data doesn't worry about locks and things because it's not directly accessing it, it's just sending a message.

The system worries about locks and things, because the message mailboxes are shared state to some extent. If you can cleanly separate ownership of different pieces of data into different actors, you get parallelism. If there's too much dependency between the pieces of data, so that separating ownership makes things more complex, maybe that becomes more visible.

That sounds nice, but it means that all reads and writes are serialized. You can't scale a system like this to, say, work for sign-up and retrieval of all Gmail users.

To scale up, you need to partition/shard the space.

For gmail users, you could hash on the (normalized) username, and all users with hash between A and B would go to one actor, etc. You don't need (or want) parallelism for signup on a single user id, but you could hash or otherwise divide that work to that level.

You can also do a partitioned work queue with a (smallish) worker pool per partition; and have the queue keep track of any keys with a current worker, to assign further work to that worker (to avoiding parallel work on the same key and the locking that requires)

These are elegant solutions, but they are also rather complex, and I'm sure that if we dig in deeper there are further complexities to explore.

I have no doubt that the actor model or continuations can be used to solve any problem. I do however doubt that it would turn out to be simpler than using shared state and explicit locking; or using ledger-like approaches. I also don't think it would be more complex!

I believe that problems of distributed computing are inherently hard enough that the incidental complexity from your chosen modeling approach is not likely to be a significant portion of the overall complexity.

The nice thing, if you do it right, is that each piece is going to be simple.

There may well be emergent complexity in how the simple parts go together, but that's going to happen anyway. Might as well make the parts simple where you can.

Incidentally, I think email is a great example of something with clear divisions of work --- an actor could easily be responsible for a single user's mailbox, and there should be no overlap in responsibility. Possibly, you could have one actor per folder, but gmail doesn't really have folders, so...

I think that's a good example of how these kinds of static allocations are tricky to get right. With your scheme, you would have hundreds of millions or even billions of actors being idle all day long, and probably a few thousand actors potentially overwhelmed with processing new email every second. You would probably have to go back and decide a new partitioning scheme to avoid this, with maybe cascading consequences in other places.

The pieces are simpler, but you might pay a greater cost in system-level design. In shared state models, you could have a more dynamic approach with a number of worker threads scaling with the number of requests (emails), with the only problem being contention when accessing the same piece of the shared storage, but at least no massive overhead from idle workers. The shared-state layer would be responsible for the locking strategy, and could even potentially optimize the required locking without necessarily changing the worker code, depending on the level of abstraction it offers workers to begin with.

We're getting into the weeds here, so don't feel the need to reply. I agree there's different complexities, and both ways will be complex in some places, and simple in others. I just want to address this one thing.

> you would have hundreds of millions or even billions of actors being idle all day long

> but at least no massive overhead from idle workers.

You could certainly work to activate the actor for a user only when it's needed (while the user is online, or while an email is delivered); but the level of overhead for an actor is quite low --- it would be fine to have the actor simply always running, and doing nothing when there's nothing to do. You can run a million Erlang processes on a server, and if they're mostly idle, there's no big deal; if you need to run a billion processes, you would need a thousand servers; but my wild guess is storage is going to be a bigger factor than CPU, assuming that the gmail style large quotas do get significantly used, I could easily be wrong about storage requirements though.

Agreed, we're getting in too many vague details.

I didn't know that Erlang can go up to such large numbers of processes, I knew they were very lightweight, but would have guessed maybe at most thousands to tens of thousands. Very nice that they can do that!

Well, you will only have to worry about mutable global data if that's what your solution calls for. Essential complexity vs accidental.

There are significantly more calling semantics beyond “push a frame and branch” as Kragen enumerates. Some, like coroutines, are visible to the high-level programmer; others, such as tail call optimization, might be invisible.

Not sure what you mean by “automatically managed” since the function preamble (before user code is run) saves registers etc. well, the abi specifies if the caller or callee does it but it’s a much of a muchness.

I mean saving the address for control to return to so that the subroutine never has to know or care; it just executes the RET instruction. And RET is just POP+JMP. This is super handy--so handy that alternative control strategies aren't explored as much as they might have been otherwise.

Reply on news.ycombinator.com