back
66 comments
Interesting talk. He mentions Futhark a few times, but fails to point out that his ideal way of programming is almost 1:1 how it would be done in Futhark.

His example is:

  sequence
    .map(|x: T0| ...: T1)
    .scan(|a: T1, b: T1| ...: T1)
    .filter(|x: T1| ...: bool)
    .flat_map(|x: T1| ...: sequence<T2>)
    .collect()
It would be written in Futhark something like this:

  sequence
    |> map (\x -> ...)
    |> scan (\x y -> ...)
    |> filter (\x -> ...)
    |> map (\x -> ...) 
    |> flatten
Also, while not exactly the algorithm Raph is looking for, here is a bracket matching function (from Pareas, which he also mentions in the talk) in Futhark: https://github.com/Snektron/pareas/blob/master/src/compiler/...

I haven't studied it in depth, but it's pretty readable.

(author here) check_brackets_bt is actually exactly the algorithm that Raph mentions
Very cool. I have seen Futhark mentioned before but I don't know anything about it.

The example you showed is very much how I think about PRQL pipelines. Syntax is slightly different but semantics are very similar.

At first I thought that PRQL doesn't have scan but actually loop fulfills the same function. I'm going to look more into comparing those.

SQL.

It is a joke, but an SQL engine can be massively parallel. You just don't know it, it just gives you what you want. And in many ways the operations resembles what you do for example in CUDA.

CUDA backend for DuckDB or Trino would be one of my go-to projects if i was laid off.

My issue with SQL is lack of composability and difficulty of debugging intermediate results.
Yes, SQL is poor.

What could be good is relational + array model. I have some ideas on https://tablam.org, and building not just the language but the optimizer in tandem I think will be very nice.

You can use SQL CTE's and/or VIEW's as a composable abstraction over queries and inspect intermediate results. The language features are there.
The standard things that someone should always say when someone brings up this problem is:

• Datalog is much, much better on these axes.

• Tutorial D is also better than SQL.

Check out https://prql-lang.org/

It solves all the warts of sql while still being true to its declarative execution. Trailing commas, from statement first and reads as a a composable pipeline, temporary variables for expressions, intuitive grouping.

is it a language problem though? it's just lack of tooling.
Even in this thread people underestimate how good e.g. DuckDB can be if you swallow its quirks. Yeah SQL has many problems, but with a slightly extended language with QoL features and seamless parallelism DuckDB is extremely productive if you want to crunch bunch of numbers in the order of minutes, hours etc (not real time).

Sometimes I have a problem, I just generate bunch of "possible solutions" with a constraint solver (e.g. Minizinc) which generates GBs of CSVs describing bunch of solutions, then let DuckDB analyze which ones are suitable, DuckDB is amazing.

More generally, the key here is that the more magic you want in the execution of your code, the more declarative you want the code to be. And SQL is pretty much the poster child declarative language out there.

Term rewriting languages probably work better at this than I would expect? It is kind of sad how little experience with that sort of thing that I have built up. And I think I'm above a large percentage of developers out there.

If you want to work in data engineering for massive datasets (many petabytes) pls hit me up!
Sorry, wrong continent :)
Raph and I also talked about this subject here: https://www.popovit.ch/interviews/raph-levien-simd The discussion covers things at a relatively basic level as we wanted it to be accessible to a wide audience. So we explain SIMD vs SIMT, predication, multiversioning, and some more.

Raph is a super nice guy and a pleasure to talk to. I'm glad we have people like him around!

There were a few languages designed specifically for parallel computing spurred by DARPA's High Productivity Computing Systems project. While Fortress is dead, Chapel is still being developed.
Those languages were not effective in practice. The kind of loop parallelism that most people focus on is the least interesting and effective kind outside of niche domains. The value was low.

Hardware architectures like Tera MTA were much more capable but almost no one could write effective code for them even though the language was vanilla C++ with a couple extra features. Then we learned how to write similar software architecture on standard CPUs. The same problem of people being bad at it remained.

The common thread in all of this is people. Humans as a group are terrible at reasoning about non-trivial parallelism. The tools almost don't matter. Reasoning effectively about parallelism involves manipulating a space that is quite evidently beyond most human cognitive abilities to reason about.

Parallelism was never about the language. Most people can't build the necessary mental model in any language.

This was, I think, the greatest strength of MapReduce. If you could write a basic program you could understand the map, combine, shuffle and reduce operations. MR and Hadoop etc. would take care of recovering from operational failures like disk or network outages by idempotencies in the workings behind the scenes, and programmers could focus on how data was being transformed, joined, serialized, etc.

To your point, we also didn't need a new language to adopt this paradigm. A library and a running system were enough (though, semantically, it did offer unique language-like characteristics).

Sure, it's a bit antiquated now that we have more sophisticated iterations for the subdomains it was most commonly used for, but it hit a kind of sweet spot between parallelism utility and complexity of knowledge or reasoning required of its users.

That's why programming languages are important for solving this problem.

The syntax and semantics should constrain the kinds of programs that are easy to write in the language to ones that the compiler can figure out how to run in parallel correctly and efficiently.

That's how you end up with something like Erlang or Elixir.

Maybe we can find better abstractions. Software transactional memory seems like a promising candidate, for example. Sawzall/Dremel and SQL seem to also be capable of expressing some interesting things. And, as RoboToaster mentions, in VHDL and Verilog, people have successfully described parallel computations containing billions of concurrent processes, and even gotten them to work properly.
iirc those were oriented more towards large HPC clusters rather than computation on single node?
Chapel, at least, aims for both. You can write loops that it will try to compile to use SIMD instructions, or even for the GPU: https://chapel-lang.org/docs/technotes/gpu.html
the distinction matters less and less. Inside the GPU there is already plenty of locality to exploit (catches, schedulers, warps). nvlink is a switch memory access network, so that already gets you some fairly large machines with multiple kinds of locality.

throwing infiniband or IP on top is really structurally more of the same.

Chapel definitely can target a single GPU.

It seems like there are two sides to this problem, both of which are hard and go hand in hand. There is the HCI problem of having abstractions are rich enough to handle problems like parsing and scheduling on the GPU. Then you need a sufficiently smart compiler problem of lowering these problems to the GPU. But of course, there's a limit to how smart a compiler can be, which loops back to your abstraction design.

Overall, it seems to be a really interesting problem!

I think a good parallel language will be the one that takes your code written with tasks and channels, understands its logic, rewrites and compiles it in the most efficient way. I don't feel that I have to write something harder than that as a pity human.
mapping from channels to SIMD seems kind of intractable, its a kind of lifting that involves looking across the producers and the consumers.

going the other direction, making channel runtimes run SIMD, is trivial

Bend comes to mind as an attempt at this: https://github.com/HigherOrderCO/Bend

Disclaimer: I did not watch the video yet

Lower-level programming language, which is either object-oriented like python or after compilation a real-time system transposition would assemble the microarchitecture to an x86 chip.
What about burla.dev ?

Or basically a generic nestable `remote_parallel_map` for python functions over lists of objects.

I haven't had a chance to fully watch the video yet / I understand it focuses on lower levels of abstraction / GPU programming. But I'd love to know how this fit's into what the speaker is looking for / what it's missing (other than obviously it not being a way to program GPU's) (also full disclosure I am a co-founder).

Was trying to remember where I recognised this name, Raph Levien is the Ghostscript and Advogato creator and helped legalize crypto https://en.wikipedia.org/wiki/Raph_Levien
Went in thinking "Have you heard of Go?"... but this turned out to be about GPU computing.
Well, they said "good" :). Go we already have, that is correct.

P.S. I'm joking, I do love Go, even though it's by no means a perfect language to write parallel applications with

Haha, yeah. Idk any other language where I practically get a free parallelization+concurrency sandwich. It's kept me coming back to Go for a decade now, despite them using a signal that prevents using it for system level libraries. They literally broke my libnss-go package years ago when they selected the signal to use to control the concurrency portion of the runtime.
VHDL?
I almost mentioned it in the talk, as an example of a language that's deployed very successfully and expresses parallelism at scale. Ultimately I didn't, as the core of what I'm talking about is control over dynamic allocation and scheduling, and that's not the strength of VHDL.
I was expecting the author to at least mention Halide https://halide-lang.org/.
https://docs.ray.io/en/latest/
The audio is weirdly messed up
Yes, sorry about that. We had tech issues, and did the best we could with the audio that was captured.
prolog?
There is the logical system that turns prolog into SQL. Something like this might be useful for nuero symbolic computing. Graph languages like cypher seem a lot more limited.
Now you are backtracking (pun intended).
ctrl-f Erlang

Nothing yet? Damn...

Erlang operates on a higher level, with much, much larger chunks. Any sensible modern system will, for instance, have some sort of execution context (thread/green thread/continuation/task/whatever) associated with a single incoming HTTP request. That's very nice, and not going anywhere.

However Erlang has very little to say about parallelization of loops, or in the levels between a single loop and a HTTP request.

Nor would it be a good base for such things; if you're worried about getting maximum parallel performance out of your CPUs you pretty much by necessity need to start from a base where single-threaded performance is already roughly optimal, such as with C, C++, or Rust. Go at the very outside, and that's already a bit of a stretch in my opinion. BEAM does not have that level of single-threaded performance. There's no point in making what BEAM does fully utilize 8 CPUs in this sort of parallel performance when all that does is get you back to where a single thread of Rust can run.

(I think this is an underappreciated aspect of trying to speed things up with multiple CPUs. There's no point straining to get 8 CPUs running in some sort of complicated perfect synchronization in your slow-ish language when you could just write the same thing in a compiled language and get it on one CPU. I particularly look at the people who think that GIL removal in Python is a big deal for performance and wonder what they're thinking... a 32-core machine parallelizing Python code perfectly, with no overhead, might still be outperformed by a single-core Go process and would almost certainly be beated by a single-core Rust process. And perfect parallelization across 32 cores is a pipe dream. Unless you've already maxed out single-core performance, you don't need complicated parallelization, you need to write in a faster language to start with.)

Yeah, i too was looking for Erlang.

The thing i would really like to see is some research on how to run the Erlang concurrency model on a GPU.

Unfortunately his microphone did not cooperate.
So he wants a good parallel language? What's the issue? I haven't had problems with concurrency, multiplexing, and promises. They've solved all the parallelism tasks I've needed to do.