Updated performance tips?

[Jonathan Goldfarb]

I've really been enjoying writing Julia code as a user, and following the language as it develops, but I have noticed that over time, previously fast code sometimes gets slower, and (impressively) previously slow code will sometimes get faster, with updates to the Julia codebase. No complaint here in general; I really appreciate the work all of the Core and package developers do, and variations in performance of different codes it to be expected.

My question is this: has anyone in the Julia community thought about updated performance tips for writing high performance code? Obviously, using the profiler, along with many of the tips at https://github.com/JuliaLang/julia/commits/master/doc/manual/performance-tips.rst still apply, but I am wondering more about general/structural ideas to keep in mind in Julia v0.4, as well as guidance on how best to take advantage of recent changes on master. I know that document hasn't been stagnant in any sense, but relatively "big in any case, I'd be happy to help make some updates in a PR if there's anything we come up with.

Best Regards,

Max

[Milan Bouchet-Valat]

Le lundi 22 février 2016 à 06:27 -0800, Jonathan Goldfarb a écrit :
> I've really been enjoying writing Julia code as a user, and following
> the language as it develops, but I have noticed that over time,
> previously fast code sometimes gets slower, and (impressively)
> previously slow code will sometimes get faster, with updates to the
> Julia codebase.

Code is not supposed to get slower with newer releases. If this
happens, please report the problem here or on GitHub (if possible with
a reproducible example). This will be very helpful to help avoiding
regressions.

> No complaint here in general; I really appreciate the work all of the
> Core and package developers do, and variations in performance of
> different codes it to be expected.
> My question is this: has anyone in the Julia community thought about
> updated performance tips for writing high performance code?
> Obviously, using the profiler, along with many of the tips
> at https://github.com/JuliaLang/julia/commits/master/doc/manual/perfo
> rmance-tips.rst still apply, but I am wondering more about
> general/structural ideas to keep in mind in Julia v0.4, as well as
> guidance on how best to take advantage of recent changes on master. I
> know that document hasn't been stagnant in any sense, but relatively
> "big in any case, I'd be happy to help make some updates in a PR if
> there's anything we come up with.

I've just skimmed through this page, and I don't think any of the
advice given there is outdated. What's new in master is that anonymous
functions (and therefore map) are now fast, but that wasn't previously
mentioned in the tips as a performance issue anyway.

The only small sentence which should likely be removed is "for example,
currently it’s not possible to infer the return type of an anonymous
function". Type inference seems to work fine now on master with
anonymous functions. I'll leave others confirm this.

Anyway, do you have any specific points in mind?


Regards

[Stefan Karpinski]

Yes, ideally code should not get slower with new releases – unfortunately, keeping track of performance regressions can be a bit of a game of whack-a-mole. Having examples of code whose speed has regressed is very helpful. Thanks to Jarrett Revels excellent work, we now have some great performance regression tracking infrastructure, but of course we always need more things to test!

[Jonathan Goldfarb]

Yes, understood about difficulty keeping track of regressions. I was originally going to send a message relating up to 2x longer test time on the same code on Travis, but it appears as though something has changed in the nightly build available to CI that now gives significantly faster builds, even though the previous poor performance had been dependable... Evidently that build is not as up-to-date as I thought. Our code is currently not open source, but should be soon after which I can share an example.

Thanks for your comments, and thanks again for your work on Julia.

-Max

[Stefan Karpinski]

I'm glad that particular slow case got faster! If you want to submit some reduced version of it as a performance test, we could still include it in our perf suite. And of course, if you find that anything else has ever slowed down, please don't hesitate to file an issue.

[Harish Kumar]

I am using Version 0.3.12 calling from python (pyjulia). I do LME fit with 2.8 M rows and 60-70 Variables. It is taking 2 hours just to model (+ data transfer time). Any tips?

      using MixedModels

      modelREML = lmm({formula}, dataset)

      reml!(modelREML,true)

      lmeModel = fit(modelREML)

      fixedDF = DataFrame(fixedEffVar = coeftable(lmeModel).rownms,estimate = coeftable(lmeModel).mat[:,1],

                     stdError = coeftable(lmeModel).mat[:,2],zVal = coeftable(lmeModel).mat[:,3])

[Mauro]

Julia 0.3.12, that's a stone-age version of Julia. You should move to 0.5!

[Harish Kumar]

Will it support Python 3.4 ? I am calling this from pyjulia interface

[Mauro]

On Sat, 2016-11-19 at 18:36, Harish Kumar <harish....@gmail.com> wrote:
> Will it support Python 3.4 ? I am calling this from pyjulia interface

https://github.com/JuliaPy/pyjulia says that it is tested against 3.5,
but it doesn't say that 3.4 is not supported. So you should try.

[Harish Kumar]

Thank you. I agree on python.. but my question was did they update the Pyjulia libraries for latest Julia version? . We tried with 0.4.3 which failed 6 months back. So we revered to 0.3.4. Or is this library remain same for all Julia versions?

Any suggestion on this?

[Mauro]

On Sat, 2016-11-19 at 20:48, Harish Kumar <harish....@gmail.com> wrote:
> Thank you. I agree on python.. but my question was did they update the
> Pyjulia libraries for latest Julia version? . We tried with 0.4.3 which
> failed 6 months back. So we revered to 0.3.4. Or is this library remain
> same for all Julia versions?
>
> Any suggestion on this?

They are testing against the latest release, i.e. 0.5:
https://github.com/JuliaPy/pyjulia/blob/master/.travis.yml

You should try and file an issue if it doesn't work. 6 months are a
long time at the current julia development pace.

[Harish Kumar]

I found the cause for this ... When i run julia 0.3.2 or 0.5 as standalone (mix model) it uses all the available cores from my server, so it was fast.

If i call Julia from Python (Pyjulia), i see only one core is busy with python process (100% cpu) and all other cores are free.  Can you help me how can i force Pyjulia/python to use available cores from my server?

Regards,

 Harish

[Mauro]

On Tue, 2016-11-22 at 14:12, Harish Kumar <harish....@gmail.com> wrote:
> I found the cause for this ... When i run julia 0.3.2 or 0.5 as standalone
> (mix model) it uses all the available cores from my server, so it was fast.
>
> If i call Julia from Python (Pyjulia), i see only one core is busy with
> python process (100% cpu) and all other cores are free. Can you help me
> how can i force Pyjulia/python to use available cores from my server?

I can't help you there. If no-one answers here try posting a new
question on discourse.julialang.org

[Douglas Bates]

On Tuesday, November 22, 2016 at 1:12:26 PM UTC-6, Harish Kumar wrote:
> I found the cause for this ... When i run julia 0.3.2 or 0.5 as standalone (mix model) it uses all the available cores from my server, so it was fast.

Fitting a linear mixed effects model only uses multiple threads for the BLAS (Basic Linear Algebra Subroutine) calls and a few LAPACK calls. In Julia v0.5 you may be able to set the number of threads for the BLAS by calling, say,

BLAS.set_num_threads(4)

(or some other number) before trying to fit a model. Be aware that increasing the number of threads doesn't always make things faster. You may need to do a bit of experimentation to determine a suitable number of threads.

Can you describe the formula you are using? If you are trying to fit a "maximal" model with large-dimensional vector-valued random effects for crossed grouping factors you should be aware that most of the time fitting such models is just a convenient way of burning up a lot of computing time.

[Harish Kumar]

I have copied my code in the below link (Julia user group), let me know if you cant access that

https://discourse.julialang.org/t/julia-call-from-python3-running-in-single-core/508/5