João Freitas

The following article explains the parallelism model in Erlang/Elixir.

Elixir is the thing I do most of my public writing and speaking about. It is my default programming language for the last 5-6 years. It suits my brain. Performs well for the kind of work I typically do. And using it I have experienced very few drawbacks. Rather than writing yet another post trying to widely summarize what I think is beneficial about the language I want to try and go a bit deeper on one particular aspect I like. Not incredibly technically deep, but unpack the concepts more thoroughly. I hope this will be a series of stand-alone posts. We will see. This time we unpack Elixir concurrency (and parallelism).

My origin story is steeped in PHP. In PHP the concurrency model is pretty much whatever your web server (mod_php) or app server (php-fpm?) decides it should be. PHP generally assumes a shared-nothing model and each request is clear of whatever work is done in the others. In this regard work can run concurrently but they can also not share anything without reaching for external files, databases, queues or similar.

Then I went to Python. Over in Python .. it depends. It is really bad at concurrency. Ordinary threading is available but enormously constrained by the GIL (Global Interpreter Lock). The best solution seems to be just running more interpreters or using Python multiprocess which I believe forks new processes. I liked Python. I didn’t enjoy trying to do performant concurrency in Python though.

Elixir builds on Erlang. Fundamentally everything I’m writing about here is thanks to the Erlang virtual machine, aka. the BEAM, which has this stuff at its core. I will write in the context of Elixir. The things that are great here are also great about Erlang. Except possibly the Task module, which is kind of neat. That could also be done with Erlang but I don’t know if anyone has.

The BEAM is made to run a ton of lightweight processes. A type of green thread if you will. They take very little time to spawn, they take very little memory. A system can have millions of them. And if you need many millions of them you can bump the default limit. They are fundamentally a different thing from OS processes and threads. The BEAM itself is scheduled by the OS but on startup the BEAM creates threads for its schedulers. One per available CPU core by default. These in turn take care of scheduling processes. This is the source of concurrency and parallelism.

We won’t go into detail on the Actor model and how that works here but that is built on top of the fundamental primitive of a process. A process runs a function, the function only has access to its own local variables and private copies of whatever arguments were passed in. This is Functional Programming so everything is immutable. A process cannot change the values in another function. This function running in a process has two cool special bits of API it can use. It can receive messages and it can send messages. If the function runs to completion the Process goes bye-bye. If the process runs in an eternal loop checking for messages and only acts if a message comes in? Then we end up talking about the Actor model again which I said we won’t do. So lets move on.

A process runs on one of the schedulers. If many processes taking a long time creates a queue on one scheduler the other schedulers can start stealing work to balance it out. If a particular process goes through a lot of function calls (internally called reductions), say 4000 for example, without completing then the scheduler will pre-empt it. That means freeze it, mid-run, and send it back to the end of the work queue. This is called pre-emptive multitasking and it is what Linux does to the OS processes. In the BEAM nothing gets to hold up the line for very long. This is how it ensures low latency which is another topic I hope to unpack at a different time.

The immutable FP nature of Elixir and Erlang are pragmatic choices that were made to achieve a bunch of goals. It predates multi-core machines by a fair bit and they were targeting a different kind of concurrency, distributed systems made up of multiple machines, a cluster. The end result was accidentally convenient for achieving incredibly concurrency and parallelism on multi-core machines. Processes do not share state, they can only communicate asynchronously through message passing. This ensures it is safe to run them at the same time. Same time as in concurrent, executing the tasks during the same period of time. Also parallel, executing the tasks at the same time. All APIs that touch files or other constrained resources, stdin/stdout and such, have mechanisms for ensuring that contention is under control. The Erlang folks have spent a lot of time getting this right.

When I write my code I generally don’t need to care about this very much. Most web requests take place inside a single process. They might communicate on through to a database connection pool which involves more processes but overall I don’t need to care about the underlying primitives. I also don’t need to think about the Node.js way of racing to IO. Node.js can be highly concurrent. CPU-bound work is fine on the BEAM. It won’t disrupt the ability of other processes to get work done, both because multiple cores can do the work but also because if I spend many CPU cycles I will be pre-empted and other work can slip by. This is all handled by the web server I use, more than by me.

If I do want to do concurrent work there is the usual async/await mechanisms, implemented in the Task module. Not as weird as in JS. They just return a Task reference which is a helpful abstraction on top of the Process ID or PID. Then you can await on it to get a result or await on multiple to get multiple results. There is also a wicked function called Task.async_stream which will take an enumerable (list, map, stream or similar) and run a Task for each entry as a lazy stream. By default it has a concurrency matching the number of available cores (just like the schedulers). This is essentially a fancy shortcut for using all your machine has to offer in the service of getting the work done for tasks that are embarrassingly concurrent. It is very fun to use when bodging together scripts that you want to go fast.

Compared to PHP I have many more tools for interacting across the work that happens in my system. Process registries and groups, PubSub mechanisms for message passing, memory-based shared state accessed through safe mechanisms (ETS for example).

Compared to Python I can do multiple things at once in simple and useful ways. In a language that is as expressive and as dynamic (in the places I care about).

Compared to Javascript I can do multiple things at once easily without worrying about CPU-bound work holding up other requests and such. I get actual immutability instead of working really hard to pretend I do.

These comparisons are of course based on my understandings and experiences with the languages. I’ve stuck with languages I have some experience with but I don’t hang out with them on a regular basis currently. If I have misrepresented any of them, do let me know.

In summary; great options for concurrency and parallelism are not an afterthought in Elixir and Erlang. They are built into the very foundation of the runtime and the languages surface it to us. This matters for performance and practicality. Doing concurrency on the BEAM is not punishing, it is not risky. It is just what you end up doing by using it.

#reads #lars wikman #erlang #elixir #beam #concurrency