This piece seems to have predicted a very active field in everyday software development since then.
What are the alternative paradigms that have actually become common use? Coroutines, async/await, that's what I hear about online but what are others? I've seen people who touted zmq-communicating-processes with standard patterns as the solution to all problems, and I'm happy not to have to maintain the results.
Have we effectively “solved” the concurrency problem, and if so what's left as an exercise for the future?
Although this paper talks initially about concurrency, you can see that really he's talking about concurrency specifically for the purpose of parallelism.
Coroutines don't solve the parallelism part, because they're concurrent but exclusive.
Async/await as implemented in JavaScript doesn't solve the parallelism part either for the same reason, and async/await as implemented in C# has exactly the same problem as threads.
There are many ideas for how to solve the problem - but I think anyone who is honest will tell you none are a perfect solution to all situations where you want parallelism or concurrency.
For example to use zmq-communicating-processes effectively you need a problem where you can divide the data or tasks cleanly a-priori. We simply don't have the mathematical understanding of how to do that to some important algorithms that people really need to run in parallel today, such as triangulation or mesh refinement.
We probably need some radical new idea, or maybe it's looking increasingly like only a mix of ideas will work.
Let's say you have a domain X that you need to triangulate. You break it along a plane, into two domains A and B, about equally large.
Imagine that you have a magical black box system that can triangulate A and B.
Would this not help you to triangulate X? I can hardly believe it wouldn't. (Again, perhaps in pathological cases yes, or if a near-optimal solution is not good enough).
Iirc, triangulation iterations can replace previous cuts. So A and B might not have any meaning in the next iteration if you triangulate properly. That's why it's okay to cut at random, because it only affects outcome indirectly.
Isn't that something you could approach with work-stealing techniques? You'd need to build the work pile as you go but that seems appropriate. Maybe I'm missing something?
The problem is: given a list of tasks that you want to solve in parallel, you cannot know ahead of time which jobs use the same data. You have to start solving them to figure that out. The problem isn't distributing the work, it's knowing when it's safe to do two things in parallel.
At risk of being that guy, the actor model (ala Erlang) is pretty good at concurrency. If you're unfamiliar, it's basically no shared state, and communication with other actors (Erlang processes) by sending asynchronous messages to the other actor's message queue.
The code for each actor is usually pretty small and easy to reason about. However, emergent behavior of the system, and ordering between messages from multiple actors can become tricky. Also, exposure to this idea long term will warp your mind :)
It depends on what is racing. If you have the same/dependent information in two (or more actors), you're going to have a coordination challenge.
So try not to do that. On the other hand, everything that happens with state within an actor is inherently non-racy, because an actor is sequential code and no other actor can mess with its state.
The actor model is a great tool, but I think it's best looked at as a low level concurrency primitive. Most of the time, folks should be working with higher level constructs in conceptually simpler control flow paradigms like call-and-return (async/await) or streams.
In Erlang, all those control flow paradigms exist in library form.
You are right about the actor model being a low-level choice, but its a choice that has to be made since the whole system revolves around allowing/disallowing shared state.
Data parallelism in CUDA, OpenGL, and other GPU APIs is doing fantastically and has for decades. (If writes are allowed, these APIs technically have the same problems as threads, but in practice they're easier to deal with since traditional mutex locks and condition variables are mostly unavailable in that environment, and the APIs force you to carefully declare the sharing semantics of your data buffers.)
Most parallel (not concurrent) problems map well to the data parallel model. Even Make is basically a data parallel API with read-only constant data, just with a more complex dependency graph.
Promises in javascript have become quite popular. Unfortunately, they're not understood very well so they aren't being used much in areas where they can improve the performance of javascript applications and instead are being used to reduce nested callbacks.
Really? At my company, promises have long since won the day. Now I'm trying to get people on async/await, which gets you like 90% of the way to the simplicity of synchronous code.
Which boils down to problems created by the POSIX implementation with condvars, mutex and semaphores. No lockless and waitfree data structures.
With threads there are also minor hidden contants: limited stack size, high cost of context switches.
And random order of evaluation.
Lockless threading semantics needs to know ownership, copy or ref and relationship to be able to fix these problems. I only know a few not well-known languages who actually did a solve these problems.
My hunch: that modern programming often requires concurrent execution of software, but that most ways in which we have to model concurrency in code are at best hard to learn, and are frequently orders of magnitude harder to learn and use.
Node.js is pretty good in this sense. Except for the very hard parts, because of node's async nature, you can introduce good amount of concurrency in your code by default, resulting in a decent amount of IO being concurrent. You have to get used to a fully-async programming model though.
It would be the mainstream languages problem then or something. Concurrency with actor model is easier to learn and use than OOP, which many people seem to be able to use.
I find a lot (most?) OOP code I read is either spaghetti (with weird object interdependencies), diffuse (waay to many classes and subclasses so following the flow of computation becomes difficult) or both. It's a great tool, but perhaps harder than widely realized.
And concurrency is even harder, especially with the ever-popular "tweak until it parses/compiles, then ship" approach, or when many people are working on the same section of code.
The Hewitt Actor approach reduces independencies dramatically, at some cost for certain algorithms, and with added clarity for many others. And it scales somewhat automatically beyond one machine which is a big win these days.
The actor model is non-deterministic and doesn't solve all concurrency problems, such as creating fine-grained or irregular data parallelism. There isn't (and it may not be possible to have) one single solution to 'the concurrency problem'.
Actor model is very deterministic, but it can model unbounded non-determinism, i.e. any concurrency problem. Including fine-grained and irregular data parallelism. It's up to the compiler to generate SIMD instructions out of it, if that's what you mean.
I've always liked section 3 of this paper, specifically the concept that "infinite interleavings" make threads executing in parallel non-deterministic and difficult to reason about. That gets to the heart of why threaded programs are so prone to heisenbugs.
"They make programs absurdly nondeterministic, and rely on programming style to constrain that nondeterminism to achieve deterministic aims."
You can't write an infinite number of test cases for all those interleavings, and it requires hard thought to suss out where any problems might lie.
This talk was about Foundation DB was brought up recently, and it's pretty amazing. I recommend watching the whole thing, but to be brief they are taming the "infinite interleavings" problem through determinism.
"Testing Distributed Systems w/ Deterministic Simulation" by Will Wilson
They wrote an interesting Actor DSL that compiles to C++ and is completely deterministic, and they torture this deterministic engine with generated test cases on a cluster every night.
I guess you could say that the whole cluster is necessarily non-deterministic, but an individual node is deterministic, given an ordering of the messages it receives.
This is just my opinion, but I've never found that part of multi-threading difficult. Interleaving doesn't matter except where resources are shared between multiple threads, and the solution is to protect the resource with a mutex.
Sometimes it's hard to tell when a resource is shared, but that has more to do with not knowing how the code works than it does with multi-threading.
> Sometimes it's hard to tell when a resource is shared, but that has more to do with not knowing how the code works than it does with multi-threading.
With respect, this sort of thing works a lot better for small codebases where you're the only one working on it. Multithreading when you can't contain the entire relevant codebase in your brain is where the real challenge is.
My favorite one is where adding debug traces causes the heisenbug to disappear because the printf() inserted a memory fence somewhere deep in the logging library.
> To offer a third analogy, a folk definition of insanity is to do the same thing over and over again and to expect the results to be different. By this definition, we in fact require that programmers of multithreaded systems be insane. Were they sane, they could not understand their programs.
I actually had this exact notion when implementing pthreads for a course. I noted to myself "Gee, I keep doing the same thing and every time I get a different result... I must be insane according to the definition"
What are the alternative paradigms that have actually become common use? Coroutines, async/await, that's what I hear about online but what are others? I've seen people who touted zmq-communicating-processes with standard patterns as the solution to all problems, and I'm happy not to have to maintain the results.
Have we effectively “solved” the concurrency problem, and if so what's left as an exercise for the future?