Hacker News Reader

Comment by wodenokoto | Hacker News Reader

Python and R do not generate high performing code. At best they generate calls to high performing code.

> At best they generate calls to high performing code.

It should be noted that this is usually sufficient. But particularly for earth scale problems it can often not be.

Julia is designed to seem to win arguments as best I can tell... If you complain about the need to break abstractions and the lack of general purpose application you're accused of not understanding. When you say it slow they say you can inline assembler, and when you say that's dumb why have a high level language then, they then say well you don't have to it is fast as is and everyone else is slow, and it just devolves into circular arguments. Abstractions exist in layers for reasons.

You can obviously provide the same abstraction with different implementations that yield different performance characteristics. Julia provides the same level of flexibility (if not more) as Python without any of the design decisions which cause Python to be so slow. I fail to see how this is a contentious point.

Yeah this is a great example of what I'm saying. You obviously don't understand Python then.

when you say Julia is slow, what are you talking about? even without any fancy tricks, normal Julia code is usually the same speed as the equivalent normal C code

Yes this is another great example... If Julia is fast then why do you need to inline assembler ?

For the same reason that C/C++ allow inline assembly? Languages come in roughly 3 speeds. Slow (e.g. python/R), mostly not slow (e.g. Java/Go), and not slow (e.g. C/Rust). If you want actually fast code (e.g. the speed of BLAS/FFTW etc) you need the combination of a not slow language, code generation, and often hand-coded assembly for the most performance critical parts.

I noticed you didn't mention Julia explicitly this time because when you outline the abstractions like this it seems silly to claim something about Julia magically solves the position and purpose of these layers. I can write a PySpark job based on a tutorial that would run circles around a single core Julia process that was designed with contradictory requirements. I just don't see how Julia gets away with claiming it solves all of this in the first page of their documentation without a ton of qualifiers... Except to say that is Julia that's what they do, they make bold claims that obfuscate what performance is and where it comes from.

To be explicit about where julia fits in here, Julia is a "not slow" language (you could make an argument for it being on the faster end of "mostly not slow" due to GC) that also has enough high level features (higher order functions, macros, memory management, general ease of use) to work as a high level language. You absolutely can write a distributed python codebase that runs faster than single core julia, but doing so will likely be harder than writing the distributed/threaded Julia code that is way faster than PySpark.

Yeah citation needed on that one, but it is a dumb hypothetical on my part to illustrate the problem. Another hypothetical, how do you get junior people to support your inline ASM? If it makes this easy to do, it makes the technical debt that much more rampant.

You seem really hung up on this inline asm thing. It's not like most julia code is just inline assembly. It's something that at a rough estimate, .5% of packages use to let you squeeze out the final drops of performance that then gets wrapped in an API that looks like normal julia code. This isn't any different from C/C++ which also in some low level code bases will have calls to compiler directives or inline assembly.

The way it is brought up in arguments about how if what you're doing in Julia is slow you can put in the ASM directly in arguments leaves me with the impression that is nonetheless a core part of the "faster than <x>" claims at least. And that's a cop out.

If I remember correctly, this came up originally in the context of comparing heavily optimized Julia code to C code that had inline assembly, in which a statement was made that Julia was obviously slower than C because the C code had hand written assembly in it. Julia, like C, sometimes needs hand-coded assembly to achieve maximal speed. 80% of a fast programming language is not having semantics that are fundamentally opposed to speed (i.e. object oriented architectures that require pointer chasing, using arbitrary precision numbers everywhere, or eval semantics that prevent interpreting rather than compiling code). Languages that don't make those kinds of mistakes are "not slow". i.e. if you write similar code in them, you will end up with similar performance to C.

Yeah I guess that illustrates how everyone is talking about different things, I just think Julia shouldn't use that to claim they are not as slow as many other things, or other vague promises of high level abilities that can somehow traverse paradigms to become the right thing to do at the low level. The fact that you've similarly called out specific caveats to the argument further illustrates how... In my opinion... It is just simply intellectually dishonest how the Julia documentation categorizes others and how promoters of the language don't really know what it means to say that Julia is fast, it clearly can be made not to be.

You realize that the specific things I called out are things julia doesn't do right? Those are the things in Java and Python and most other high level languages that prevent them from being fast. Julia's semantics specifically don't do those things.

Honestly I don't know why you're replying, it is not really addressing what I'm talking about and using a lot of the argument styles I'm complaining about.

I used to think so, but I have a function that gets called about a billion times each and every day as new data comes in, and and takes about 0.01 seconds to evaluate (optimizaiton with nlopt). I tried to code it in c (30% speed improvement) python (twice as slow), Julia (about the same speed). Reason is that call has 5 parameters that operate on a vector of length 50 to return a value to minimize. Turns out R is pretty good at such vector calculations.

Is this what you mean by nlopt? https://github.com/stevengj/nlopt

If so, it looks like you're interfacing from R to high-performing code written in C. Isn't that exactly what OP was describing?

no, the function it calls is pure R and that is where the the code spends all its time.

Interesting. Have you published the code and/or benchmarks anywhere? This flies against everything I've read about Julia and R.

No, its 4 lines of code. I just benchmarked for myself. All it is for 2 vectors x and y of average length 50, and 5 parameters, with exponential, addition, and multiplication, and ultimately sum to return to the optimizer. I was also surprised as I expected c to be much faster. And with Rccp, its actually slower than the R, overhead I guess. When I looked into it, apparently R has really fast code for such vector calculations. With julia, admittingly I did not use simd which would likely make it faster.

Now, I generally use Julia for heavy computes, and usually its much faster than R. But not always.

And this little bit of code runs for hours on the largest instance on AWS every day. Why I was looking so speed it up.

If possible I would encourage you to make an MWE and post it on the Julia Discourse - "Julia slower than R for optimization problem" or something like that, there's a good chance that the community will be able to eke out some more performance. Alternatively you might have hit on a case in which Julia itself is currently leaving performance on the table, which would still be helpful for the community to know as being slow is often considered a bug in Julia world.

Great idea, I do lurk there, just might do that next time I look at the code. My hunch is that simd is the low hanging fruit julia brings. But, and I am not an expert, both might end up doing BLAS anyways, which is why they are so similar.

I am willing to concede, but also willing to argue that vector and data frame manipulations in R are calls to optimized code.

Like, R is “what if we made a lisp inspired version of Python built around numpy and pandas and then reversed timed”

I think that is exactly what is happening. Most of my code is much much faster in Julia, and the code is nicer. But R has its moments. Which is good since this particular app has 3K lines, and I do not want to port it to Julia.

And data.tables in R is faster (and I think nicer to write) than DataFrames in Julia. And since data.tables feed my optimization, R still wins.

Reply on news.ycombinator.com