Comparison of calculation speed

[pucicu]

I did some comparison of the calculation speed for RPs and simple RQA. More details at https://github.com/pucicu/RP_Speed_Test. The Julia implementation (RecurrenceAnalysis.jl) is (surprising to me) not the fastest one.

Have fun! 😁

Norbert

[Datseris]

Hi Norbert!

Thanks a lot for your post. Generally speaking I would be happy to see that there can be improvements to the library. However, are we sure that these benchmarks are accurate? I had a look at the repo you cite.

• For Pyunicorn it seems to me that the function being benchmarked is: rqa_summary. The docstring states that it computes "the recurrence rate RR, the determinism DET, the average diagonal line length L and the laminarity LAM".
• For RecurrenceAnalysis.jl, the function being benchmarked is rqa which computes:
  
  So, it computes at least triple the amount of quantities as rqa_summary, and the quantities computed are also much more complex than average rates.
• (I have no comments on the GPU version, as a direct performance comparison between GPU and CPU code can only yield the obvious result "GPU is faster", and cannot really be of any help to the CPU code. Of course, it is great that a GPU version exists for RQA!)

It comes to no surprise for me that our version is slower.

Furthermore, the benchmark also mixes the time it takes to compute the recurrence matrix with the time it takes to compute RQA measures. It would be more useful to separate those two, and would also help each library see where they can improve.

[heliosdrm]

For Pyunicorn it seems to me that the function being benchmarked is: rqa_summary. The docstring states that it computes "the recurrence rate RR, the determinism DET, the average diagonal line length L and the laminarity LAM".

Maybe it would be closer for Julia if rqa is called with the keyword argument onlydiagonal=true. As noted in the documentation:

The keyword argument `onlydiagonal` (`false` by default) can be set to `true`
in order to restrict the analysis to the recurrence rate and the parameters related
to diagonal structures (`:RR`, `:DET`, `:L`, `:Lmax`, `:DIV` and `:ENTR`), which makes
this function slightly faster.

Also, I have not checked if Pyunicorn or other alternatives run in parallel threads by default. This package does not, unless parallel=true is passed to the calculation of recurrence matrices.

[pucicu]

I know that some of the tools I compared are not computing all measures. But pyunicorn computes measures quantifying the diagonal and the vertical line structures, thus, requires finding the histograms of them. The calculation of the histograms is the most time consuming part, the different quantifiers, e.g., LAM, TT, Vmax, Ventr etc. are easy to calculate and do not need so much time as the histogram itself. You can only complain about the missing "white vertical" lines.

Surely, for a deeper test we should separate the calculations of RP and RQA. For the purpose, for which I implemented this test, such a summary was completely fine. And also in your comparison I found such a combination of calculation times: https://github.com/JuliaDynamics/RecurrenceAnalysis.jl/wiki/Comparison-of-software-packages-for-RQA.

[Datseris]

I was not aware of the cited page https://github.com/JuliaDynamics/RecurrenceAnalysis.jl/wiki/Comparison-of-software-packages-for-RQA before. I only learned of it now. I have not even read it yet (haven't had the time yet!!!)

I've already opened a PR to add it to the official docs: JuliaDynamics/DynamicalSystems.jl#172 . In my opinion, documentation should be in one, and only one place. Mixing github wikis and readmes and offiicial docs makes information harder to locate, and much harder to maintain as well. So, we will update the page accordingly and put it in the official docs so that it is also more discoverable!

In the same PR we will also add a new comparison of performances!

The calculation of the histograms is the most time consuming part, the different quantifiers, e.g., LAM, TT, Vmax, Ventr etc. are easy to calculate and do not need so much time as the histogram itself. You can only complain about the missing "white vertical" lines.

Okay, yes, that is a fair point.

[Datseris]

@pucicu does rqa_summary run on parallel threads by default?

[pucicu]

@pucicu does rqa_summary run on parallel threads by default?

I'm not sure, but I don't think so. At least the CPU load is not showing such. MATLAB is running parallel.

Maybe it would be closer for Julia if rqa is called with the keyword argument onlydiagonal=true. … Also, I have not checked if Pyunicorn or other alternatives run in parallel threads by default. This package does not, unless parallel=true is passed to the calculation of recurrence matrices.

I have just tested both but could not find a big improvement, only a small one.

[pucicu]

I have tested the calculation time for RP and RQA separately. The result is really interesting.

I could not include the R package, because we cannot separate the calculations there. In the MATLAB version, I have commented out the network measures. But I have also included a version using loops. Now MATLAB (vector/ pdist) has the same performance as pyunicorn for RP, and same as Julia for RQA.

But I'm quite sure that I did some stupid mistake with the Julia implementation of the loop and it might be possible that it could be much faster. And I could not find a difference in Julia using parallel=true or false and onlydiagonal=true or false.

[Datseris]

Thanks a lot Norbert. This is very helpful.

. And I could not find a difference in Julia using parallel=true or false and onlydiagonal=true or false.

Okay, that is clearly a problem on our end, as these settings should be making a difference. For the parallel setting to work you would need to start Julia with multiple threads, which doesn't happen by default. I think either me or @heliosdrm should take a look at your repo when we have time, and maybe open Pull Requests for the Julia code snippets, if we think this is necessary.

Given that the pyunicorn version is clearly faster, we should also look at its source code and see what techniques we can copy from there ;)

[pucicu]

Hehe, I have now checked the results and found them a bit strange. I had used the z-component of the Rössler as Horatio did in his comparison. But this causes quite large number of recurrences (RR=0.56). Now I switched to the x-component which gives better RPs and much lower recurrence density (RR=0.03). Now the speed is much better for Julia. Nevertheless, it seems that pyunicorn is more efficient for higher RR than Julia.

And it means that all the results above hold only for larger RR.

[Datseris]

Didn't we had a discussion somewhere, with a PR that @hkraemer started, and we made a choice regarding this? We had two versions to compute something and one was better for one recurrence and the other...? I don't remember very well, please correct me if I am wrong.

[pucicu]

I also remember only vague. Hauke should know.

[heliosdrm]

Yes, it was about the skeletonization algorithm (#128 ). It resulted that the algorithm that is used to make the histograms works better the more sparse are the matrices. This is consistent with what Norbert says.

[pucicu]

but does it be included per default?