I’ve always been surprised that we don’t have a really widely supported construct in programming that is like a for loop, but with no dependency allowed between iterations. It would be convenient for stuff like multi-core parallelism… and also for stuff like out of order execution!
Not sure how “break” would be interpreted in this context. Maybe it should make the program crash, or it could be equivalent to “continue” (in the programming model, all of the iterations would be happening in parallel anyway).
I vaguely feel like “for” would actually have been the best English word for this construct, if we stripped out the existing programming context. I mean, if somebody post gives you instructions like:
For each postcard, sign your name and put it in an envelope
You don’t expect there to be any non-trivial dependencies between iterations, right? Although, we don’t often give each other complex programs in English, so maybe the opportunity for non-trivial dependencies just doesn’t really arise anyway…
In math, usually when you encounter “for,” it is being applied to a whole set of things without any loop dependency implied (for all x in X, x has some property). But maybe that’s just an artifact of there being less of a procedural bias in math…
We actually do have the abstractions but the problem is that the vast majority of for loops don’t benefit - you need to have so much work that the overhead of coordinating the threads is sufficient. Additionally, you’ve got all sorts of secondary effects like cache write contention that will fight any win you try to extract out of for loops parallelism. What we’ve been learning for a long time as an industry is that you benefit most from task level parallelism with minimal to no synchronization.
Granted this probably isn't the parallel application that the other poster was envisioning, but it can be extremely useful when a computation depends on a large number of I/O-bound tasks that may fail, like when you are servicing a request with a high fan-out to other services, and you need to respond in a fixed time with the best information you have.
For example, if you need to respond to a request in 100ms and it depends on 100 service calls, you can make 100 calls with a 80ms timeout; get 90 quick responses, including two transient errors, and immediately retry the errors; get eight more successful responses and two timeouts; and then send the response within the SLA using the 98 responses you received.
That doesn't require parallelism, just concurrency. But yes, you'd use a similar task-local map/reduce construct to express doing a bunch of concurrent I/O in parallel (spawning each I/O on a separate thread would be counter-productive & a hack to enable not adding an event loop / async I/O).
I’m only familiar with Fortran and my openMP is a little rusty. But, I think there are different pragmas for vectorization or threading. So, you have to tell it to do one or the other (is that wrong). Instead of expressing “well we don’t have dependencies here so do what you will.”
That would be great if the engine or compiler had that kind of capability but building that requires solving the halting problem.
Even if you try to do it with heuristics, go ask Itanium how that worked out for them and they tried a much simpler problem than what you’re proposing.
I am not familiar with the Itanium story and I don't know who to ask.
But it seems to me like this would be a safe space to experiment. With heuristics and pragmas as a fallback. Because with the right approach solutions would mostly be better than not doing anything.
And you could do it in runtime when you know the size of the input.
And what about applying the logic to places where you can see that the loop will end?
I believe query planners in for example Trino/BigQuery do this already?
> like a for loop, but with no dependency allowed between iterations
"Break" is a dependency between iterations, and really only makes sense in a sequential iteration. In a parallel for loop, you can break from the current iteration, but the next is probably already running.
If you want any iteration to be able to cancel all others, they have to be linked somehow. Giving every task a shared cancellation token might be simplest. Or you turn your for loop into a sort of task pool that intelligently herds threads in the background and can consume and relay cancellation requests.
But I agree, we need a new paradigm for parallel programming. For loops just don't cut it, despite being one of the most natrual-feeling programming concepts.
C#'s Parallel.For and ForEach are a step in the right direction, but very unergonomic and unintuitive. I think we could get by with just bolting parallelism onto for loops, but we need a fundamentally parallel concept. I assume it'd look something like cuda programming but I really don't know.
The tricky cases are the very many where there are dependencies between iterations, but not demanding the strict serialization that a simple loop enforces. We have constructs for that, but there's an irreducible complexity to using them correctly.
They're not in the language proper, but "parallel for" is a common construct. I've seen it in C# and Rust, but I'm sure other languages have it too.
It may be a good idea to use a framework with explicitly stateless "tasks" and an orchestrator (parallel, distributed, or both). This is what Spark, Tensorflow, Beam and others do. Those will have a "parallel for" as well, but now in addition to threads you can use remote computers as well with a configuration change.
The big C and Fortran compilers have openMP support, which includes parallel for loops. They just feel kind of… bolted on, being a pragma based language extension. And what I really want to express to the thing isn’t “fork here” but “here are some independent operations, tell the optimizing compiler about it,” and then the optimizing compiler can (among other transformations also decide to sprinkle some threads in there)
> we don’t have a really widely supported construct in programming that is like a for loop, but with no dependency allowed between iterations
Uhhh... we don't? It seems to me like we do. This is a solved problem. Depending on what you're trying to do, there's map, reduce, comprehensions, etc.
And for those who also don't want to be forced to sequence the computations, i.e., wanting to run them concurrently and potentially in parallel, each approach to concurrency supports its own version of this.
For example, choosing Scala on the JVM because that's what I know best, the language provides a rich set of maps, folds, etc., and the major libraries for different approaches to concurrency (futures, actors, effect systems) all provide ways to transform a collection of computations into a collection of concurrent operations.
Curious if the poster who said "we don't have a really widely supported construct" works in a language that lacks a rich concurrency ecosystem or if they want support baked into their language.
Sounds like you're talking about Realtime operating systems. I don't know if there are many/any programming languages that build those operational requirements into the syntax/abstraction.
For loops are the "goto":s of the parallel programming era.
Ditch them and the rest can be handled by the programming language abstraction.
Why? Because they 1. Enforce order of execution and 2. Allow breaking computation after a certain number of iterations.