Hi all, We wrote a new blog post about comparing Julia with MATLAB, Python, and R. Comments and suggestions are welcome! [image] Choosing a numerical programming language for economic research: Julia,... Economic researchers can choose between four general purpose programming languages – Julia, MATLAB, Python, and R. This column evaluates the four, comparing and contrasting the various language features and making recommendations in particular use...

[image] DNF: @turbo is single-threaded, while @tturbo is multi-threaded. The extra t stands for threading. Yes, of course. Still it uses SIMD and we could have used that for (some of) the other languages, so doesn’t quite make for a fair comparison. But certainly worth mentioning.

I disagree. Automatic simd is definitely fair, since it doesn’t use extra resources, just exploits the allocated ones better. In fact, this is not at all limited to @turbo, but is something Julia uses automatically in many different places. Would you block simd-vectorization somehow deliberately? Anyway, numpy and Matlab also surely uses simd internally, probably R too. Limiting multi-threading makes sense for comparisons, but preventing simd does not make sense.

[image] TheCedarPrince: Inspired by @Sukera 's comment a bit up the chain about test documentation and @johnmyleswhite 's comments on identifying and trying to tackle issues, I opened a PR to the Julia docs that gives an example workflow on how to use pkg> test within one’s own package going through how to create tests, run the tests, and add the test suite to one’s own package. A fantastic idea. some time my misunderstanding of Pkg let me to initiate a discussion of here, where I showed myself to be rather ignorant of Julia, alas, I have no shame and here it is, if of use. Best practice for packages on shared drives

Thanks and no worries - it’s discussions like these that can lead us to solutions. :smiley: I figure even if that isn’t a perfect PR there, it is at least something now we can all critique and make better. Feel free to comment on that PR with additional ideas/thoughts.

I would agree on most cases. However, I believe the authors specifically used an algorithm that can’t be vectorised. The improved Julia code uses a modified algorithm which can be vectorised in part. It’s not really a fair comparison in the spirit of the article.

Agreed, and @turbo is no less fair than @jit.

In case you are interested I also made a functional version of the code that maybe can be parallelized. It is not faster, but I like it because it has zero allocations and the GC is not called. using StatsBase using DelimitedFiles using BenchmarkTools using Base: Iterators using IterTools y = readdlm("data.dat") y = y[:, 1] N = length(y) S = 1e3 o, a, b = [0.001, 0.85, 0.01] y2 = y .^ 2; v = var(y); function likelihood_functional(o, a, b, h0, y2) N = length(y2) function updateh(t) newh, i = t @inbounds o + a * y2[i] + b * newh, i + 1 end function loginv(t) hh, i = t @inbounds (i>1) * (log(hh) + y2[i] / hh) end return mapreduce(loginv, +, Iterators.take(iterated(updateh, (h0, 1, false)), N)) end lik = likelihood_functional(o, a, b, v, y2) bench = @benchmark likelihood_functional($o, $a, $b, $v, $y2) println("output,Julia-functional ", VERSION, ",", lik, ",", minimum(bench.times)) It is not really faster, but I like about this is that the timings are more consistent BenchmarkTools.Trial: 10000 samples with 1 evaluation. Range (min … max): 92.477 μs … 126.451 μs ┊ GC (min … max): 0.00% … 0.00% Time (median): 94.428 μs ┊ GC (median): 0.00% Time (mean ± σ): 94.960 μs ± 2.134 μs ┊ GC (mean ± σ): 0.00% ± 0.00% ▅██▄▇▇▅▂▂▁▂▂ ▂▁▁▂▂▃▁▁▁ ▁ ▂ ▆▄▃▇▅▄▇▆█████████████▇██████████▇▅▆▇▆▆▆▆▇▆▇▆▇▅▄▅▅▄▄▅▄▄▅▅▄▇██ █ 92.5 μs Histogram: log(frequency) by time 102 μs <

[image] DNF: I disagree. The way we wrote the code in all 5 languages does not use SIMD so the comparison is not fair. Assuming one can do so (which we have not looked at) we would be happy to include a separate comparison between all 5 languages that use SIMD. For now, we will add the SIMD Julia code and timing to the appendix. We already changed the appendix to include @turbo. We would love to credit @Jorge_Vieyra but want his approval first.

[image] Jorge_Vieyra: For the Julia code this is a faster version: Jon was impressed with this proposal, he runs it every day on his website http://extremerisk.org/ (the estimation is done in Julia) and it will significantly speed up his code. And be more elegant.

Choosing a numerical programming language for economic research: Julia,

Community

DNF August 15, 2022, 6:12pm 82

Changing the algorithm might not be fair game, I agree, but any automatic simd that might happen should be fine.

Topic		Replies	Views
Julia packages that have no equivalents in other languages? Teaching & Outreach	42	7534	July 30, 2021
Finance and Economics Use Cases Finance and Economics	105	18334	March 11, 2024
Why Julia is very often related to a data science / economics language only? Offtopic	46	4010	April 22, 2022
A Comparison of Programming Languages in Economics Community	62	7352	August 7, 2018
Things that are easier in Julia than Python/R etc Community python , r	59	8143	October 17, 2021

Choosing a numerical programming language for economic research: Julia,

Related topics