Hacker News Reader

Comment by eigenspace | Hacker News Reader

Version 1.5 is an awesome release with some really nice stuff. That said, I can't help but feel frustrated by all the attention and effort being poured into compiler latency. I get that some people value it very highly, but to me it just seems silly to put so much effort towards this big group of whiners who find it unacceptable to wait __ seconds for the compiler to run before their first plot is produced (especially when PackageCompiler.jl exists!). To me, the important thing about julia is runtime performance. I'd gladly pay hefty compile times to have even minuscule runtime improvements.

To be clear, I'm not saying the Julia devs have been ignoring runtime performance, far from it, but I just sometimes feel the need to chime in like this because I often feel the compiler latency complainers are disproportionately squeaky wheels, so I'd hate to see that cause the devs to over-correct.

View on HN

I don't really want to sacrifice runtime performance for compile time performance, so long as the compilation time is "reasonable", but so far we've mostly been having our cake and eating it too. The eliminating-invalidations work for Julia 1.6 is a great example of that.

When it comes to explicit trade-offs, like setting the optimization level, they've given us finer control by letting us specify it at the module level. Now Plots.jl can favor compile time, while I'll numerical libraries can favor runtime.

Right, so what I'd say is that while latency improvements and runtime improvements to julia aren't necessarily in conflict, the core developer's time is a zero sum competition between these things (among others).

I'm just saying that for selfish reasons, I'd rather they spent their time on runtime and infrastructure improvements than reducing compile times.

And of course, I say this while acknowledging that I am in no position to demand how they spend their time.

Sometimes aspects that help the community are a better long term investment, and I think that most people would point to compile times as the key thing to nail down in order to grow the community at a faster rate.

But the runtime performance is not that great either.

  Dot product:      NumPy 0.03528 vs 0.03 Julia    (x1/1.1)
  Element-wise sum: NumPy 0.00379 vs 0.0063 Julia  (x1.6)
  Element-wise mul: NumPy 0.00419 vs 0.00617 Julia (x1.5)
  L2 norm:          NumPy 0.02391 vs 0.097 Julia   (x4.1)
  Matrix product:   NumPy 0.00186 vs 0.01988 Julia (x10.7)
  Matrix sort:      NumPy 0.01033 vs 0.0161 Julia  (x1.6)

I tried element-wise sum, and got

  julia> using LoopVectorization, BenchmarkTools

  julia> A = rand(10_000, 10_000); B = rand(10_000, 10_000); C = similar(A);

  julia> @benchmark vmapntt!(+, $C, $A, $B)
   BenchmarkTools.Trial:
    memory estimate:  13.19 KiB
    allocs estimate:  91
    --------------
    minimum time:     29.103 ms (0.00% GC)
    median time:      29.202 ms (0.00% GC)
    mean time:        29.329 ms (0.00% GC)
    maximum time:     30.658 ms (0.00% GC)
    --------------
    samples:          171
    evals/sample:     1

While with NumPy

  >>> import timeit
  >>> u = timeit.Timer("A + B", setup='import numpy as np; A = np.random.rand(10_000, 10_000); B = np.random.rand(10_000,10_000)')
 
 >>> u.repeat(10, 1)
  [0.17918606100283796, 0.17888473700440954, 0.17893354399711825, 0.1790916720055975, 0.17922663199715316, 0.17935074500564951, 0.17939399600436445, 0.17940548500337172, 0.17933111900492804, 0.17920518800383434]

A 6x advantage for Julia. If I don't use multithreading, Julia slows down to 122.5ms, or about 1.5x faster.

Elementwise multiplication is the same fast.

If I get a little more creative and do exp(A) + log(B) instead, multithreaded Julia still takes 30ms, while single threaded slows down to 169ms.

Meanwhile, NumPy slowed down to 847ms, making it 30x slower than multithreaded Julia (on my computer) and 5x slower than single threaded.

Shall I keep making programs more complicated?

Never use big arrays in benchmarks otherwise depending on your CPU cache you will be testing allocation as well. If you need the code I tested Julia on here you go: https://github.com/tastyminerals/mir_benchmarks/blob/master/...

Thanks for sharing the benchmarks.

I made the arrays huge to get roughly the same ball park of times you reported. I also preallocated the memory in that benchmark to avoid unnecessary allocations (the observed allocations are because multithreading in Julia allocates memory). Is this something as easy to do in Python?

I made a few tweaks to your benchmark code. I get with Julia:

  Element-wise sum of two 100x100 matrices (int), (1000 loops)
    1.698 ms (2 allocations: 78.20 KiB)
  Element-wise multiplication of two 100x100 matrices (float64), (1000 loops)
    1.786 ms (2 allocations: 78.20 KiB)
  Dot (scalar) product of two 300000 arrays (float64), (1000 loops)
    4.801 ms (0 allocations: 0 bytes)
  Matrix product of 500x600 and 600x500 matrices (float64)
    229.475 μs (0 allocations: 0 bytes)
  L2 norm of 500x600 matrix (float64), (1000 loops)
    6.566 ms (0 allocations: 0 bytes)
  Sort of 500x600 matrix (float64)
    523.243 μs (595 allocations: 2.32 MiB)

Python:

  | Element-wise sum of two 100x100 matrices (int), (1000 loops) | 0.0042680892984208185 |
  | Element-wise multiplication of two 100x100 matrices (float64), (1000 loops) | 0.0027007726486772297 |
  | Dot (scalar) product of two 300000 arrays (float64), (1000 loops) | 0.0088505959516624 |
  | Matrix product of 500x600 and 600x500 matrices (float64) | 0.00107754109922098 |
  | L2 norm of 500x600 matrix (float64), (1000 loops) | 0.010691180499998154 |
  | Sort of 500x600 matrix (float64) | 0.008244920199649642 |

That is, Julia-speedup factors of:

  1. Small int matrix element-wise sum: 2.5x
  2. Small float matrix element-wise product: 1.5x
  3. Dot product (scalar): 1.84x
  4. Matrix multiplication: 4.7x
  5. L2 norm: 1.6x
  6. Sorting: 15.8x

The changes were: 1. Memory allocation was slow, but Julia makes in-place operations easy, so I made 1, 2, and 4 in-place. Although I still allocate the result within the test function for `1.` and `2.`. 2. MKL is much faster than OpenBLAS, so I made `BLAS.vendor()` return `:mkl` 3. I sorted in a threaded loop and transposed the destination array. Transposing the destination array was because Julia is column major, so sorting rows is a benchmark that will naturally favor the row-major language.

But I left all the inputs the same.

My primary point with Julia here is that it's easy to optimize it and get more performance out of it if or whenever you need it.

Now try doing these tests on arrays of user defined types. Or use one of those fast functions as the kernel of an ODE solver and compare to julia. Anyone can game microbenchmarks on arrays of Float64s, doing so shouldn't impress anyone. If Numpy can do it, you know it's not impressive.

The point of Julia is that Float64 and Int aren't special. They're on the same footing as whatever types you cook up, and our performant multiple dispatch system lets you build a vibrant interacting package ecosystem filled with user defined types doing targeted overloads of other functions.

It's the multiple dispatch and composability of julia code that matters. The important part of the performance of Julia is that we can get the ecosystem benefits of writing pure julia code without paying with runtime performance penalties.

Even for floats & ints, try doing some array operations without allocating memory for intermediate results. In Julia, I can write whole PDE simulations that never allocate intermediate values inside the time integration loop. I haven’t done serious work with NumPy in a while, but I’m pretty sure that’s impossible for anything non-trivial, even a simple FD stencil such as `du[2:end-1] = D * (u[1:end-2] - 2*u[2:end-1] + u[3:end]) / dx^2`.

Absolutely. Numpy's performance model is so brittle that one needs to be very selective when they cherry-pick their microbenchmarks.

My rule of thumb is that if your cherrypicked microbenchmark involves one specialized Numpy function call, it'll be fast, but if there is two separate specialized Numpy function calls it'll be slow.

Aren't many of those benchmarks mostly comparing MKL vs OpenBLAS (the default BLAS library used)?

Also, it's unlikely that Julia will handily beat the handwritten, carefully optimized C-code in numpy. At that level, it isn't very much about the language but more about the programmers implementing the algorithms. A major selling point of Julia is that you will have fast code even in cases where the code doesn't neatly factor into calling static kernels written in e.g. C.

But this is also true for any other statically typed language. Java/Scala/Kotlin/Go/Rust you name it. Pure Java has a good chance of outperforming a mix of Python and NumPy. So what's Julia's selling point? Being dynamic and fast? But it comes with some really nasty side effects. Up to the point that it is sometimes doesn't even make sense and Go compile & execute combo doesn't feel much slower.

At this stage I can only see that it is the ecosystem which is hard to evaluate without investing into learning the language. However, I find it hard to believe that it is richer than Java or Python.

> But this is also true for any other statically typed language. Java/Scala/Kotlin/Go/Rust you name it.

Notice how none of those languages have a vibrant numerical computing ecosystem anywhere near Julia's despite being just as old or older than julia and being big, well known languages with large communities? It's almost like scientific programmers don't want to do science in those languages.

I don't think it's an accident that Julia's package developers have been remarkably productive compared to Python. The best differnetial equations library available on Python right now is a wrapper for Julia's DifferentialEquations.jl. This is not because Python users don't care about differential equations or because there's not enough money or manpower. Chris Rackauckas did more on his own in julia for differential equation solving in his first couple years than entire industries did in Python over decades.

Julia is a productive and performant language. When you hold up Java and Python, you're asking me to choose between productivity and performance (at best, obviously Python has serious productivity problems and Java has serious performance problems despite those being their touted strengths).

Come on. It's not like Julia has a PyTorch / TF equivalent (Flux.jl doesn't even come close, sorry). Until then, you can't seriously say that Julia's ecosystem is rich. Once we start seeing DL/ML papers coming out with POC written in Julia we can call it a day.

So you're disappointed that Mike Innes and a handful of other open source contributors haven't quite caught up yet with the efforts of Google and Facebook pouring billions of dollars into TensorFlow and PyTorch in pursuit of one of the most lucrative and hyped modern markets?

Well, for that it depends on what you're into. For traditional, coding bootcamp machine learning, Flux.jl is still behind, but I'd point out that the only high performance, usable NeuralODE library is still built on Flux.jl and DifferentialEquations.jl as far as I understand and mostly consists of some import statements and a few function compositions / overloads. Someone already did deep learning on fully homomorphic encrypteded data using Flux.jl. Many 'niche', cutting edge things are still best done with Flux.jl because Flux is fundamentally more flexible than things like PyTorch and TensorFlow. It's using the entire programming language as a framework in a very generic way that PyTorch and Tensorflow can only ape by building ever more complex custom compilers and DSLs with ever diverging semantics from Python.

If Flux.jl is an example of the Julia community's failures to compete I'd say we have very little to worry about, and I look forward to seeing what a well funded commercial entity could build in Julia with a few billion dollars.

Consider me a curious client who trains and deploys production models on a weekly basis and considers Julia. Well even if I could rewrite our model in Julia how would I go with prod deployment and who would help me if I get stuck? For Tensorflow, I can use Java API and effortlessly plug in trained model into our Java stack. In case of PyTorch we can bring up a Python service with the same little effort just because Python has it all.

When it comes to work, niche languages lose their ground once you start digging deeper and attempt to venture slightly outside their focus domain. They do not have dedicated teams or industrial money and will always lag behind. Look at DL4J, they have a whole company working on it for many years.

Of course time will tell but considering the results so far I am not convinced that investing time into learning a new language will pay off. Especially considering Julia interop with C/C++.

> Especially considering Julia interop with C/C++.

Is the interop bad?

As far as I'm aware, Julia interop with C/C++ is particularly excellent. So, maybe the poster can clarify?

Well, it's been 4 years so maybe it's better now: https://www.zverovich.net/2016/05/13/giving-up-on-julia.html

Well, running "Hello, World" still takes 0.3 ms :D

That is truly the dumbest critique of Julia that's been published. So dumb I didn't bother refuting it at the time, but apparently people take it seriously?

- Hello world is slow? Really? Yeah, it's not a hello-world-oriented programming language.

- One-based indexing is questionable? Troll much?

- Using `ccall` is unsafe? No kidding, there's no way to call a shared library that doesn't crash if you lie to it.

- `@printf` sucks? Then don't use it. I never do. The `print` function works for 99% of cases.

- Julia development is slowing down? Whiffed on that call.

Regarding the C ffi point, you can use Clang.jl to generate correct bindings so that you don't have to read C header files. But note that if you want to call a function in C you need to look at and match its signature, so it's not all that onerous to have to do the same from Julia. Julia's C interop is so good that several other languages have copied it since it came out.

Yeah, but how else would you judge about the language before spending months on learning it? Critique can be dumb like 1-based index for arrays or nitpicking on syntax. However, I tried Julia several years ago and start-up times was the reason I dropped it. Was it a dumb thing to do too? I was disappointed that a promise of being a dynamic language came with some "but"s. And if you have a lot of scripts that you need to run one-by-one, you don't notice the real speed of the language, all you notice is sluggishness.

If one isn't willing to spend some time on learning a language, one simply shouldn't write a scathing blogpost about it.

> So dumb I didn't bother refuting it at the time, but apparently people take it seriously?

You definitely should have because from what I saw during my google searches this idiotic article was mentioned a couple of times.

You aren't making a fair comparison here.

Flux is more generic, flexible and thus harder to build. Julia could have been there already if the small ML team w was working with TF type design goals. But they aren't, and therefore are ahead of TF/pytorch in many respects.

Also not as much effort has gone into making this polished for standard deep learning stuff, but you can see the fruit in things like reusing the same code for ML on holomophically encrypted data, faster than pytorch/tf deep learning on graphs and 3d inputs, faster, easier composability with ODEs and prob programming etc.

A more fair comparison is KNet.jl which is a very worthy competitor to things like TensorFlow or PyTorch with comparable speed and roughly equivalent flexibility. It's just not talked about as much by Julia people because what most Julia people really care about is the flexibility and composability of Flux, so they're willing to forgive it lagging a bit in raw performance (as long as they believe it'll catch up eventually).

Neither does Python, PyTorch / TF equivalents are available to any language with FFI to C and C++ libraries, which are the languages they are actually written on.

Interfacing with C++ is a lot of work and there is no language in existence that can do that effortlessly. Nim and D are the ones that come close. Julia's interoperability with C/C++ is quite poor according to feedback from more experienced Julia users.

.NET does it relativity easy given C++/CLI, and on the native side there is COM, UWP, C++/CX and C++/WinRT.

So any language on Windows able to use COM/UWP, gets those Python libraries, sorry, C++ libraries.

The effort to use Julia FFI to interoperate with them is hardly any different than using Python FFI, with the big difference that in what concerns Julia it would be possible to re-implement them in pure Julia, while the Python community keeps wondering about to write more FFI code to native libraries or finally start adopting PyPy.

This makes no sense without showing what you're running ? Are you also using MKL in Julia? https://github.com/JuliaComputing/MKL.jl

Numpy is written in C. What are Julia’s functions written in?

Julia was written in C++ in early stages if I remember correctly then Julia ;) The point here is that NumPy is heavily fine-tuned and will beat almost any other language implementation out there. The community is bigger and therefore there are much more free hands.

In Julia's case, there are maybe more "free hands" who can contribute to performance critical code. To contribute to `NumPy` you have to know how to program in C.

One of the important facts about OSS development is that the number of users doesn't scale linearly with the number of "free hands. There are less than 2 FTE's working on Python at any given time these days.

Reply on news.ycombinator.com