Interesting fact. It seems like using newer Oracle...

# mathematics

[altavir]

Interesting fact. It seems like using newer Oracle GraalVM JIT significantly improves performance even for Multik

[altavir]

I see about 3x performance imporvement for EJML and about 2x improvement for Multik in the same benchmarks.

[altavir]

Another fun fact. Switching two lines in KMath (turning on parallel buffer processing) allows to rump up performance almost to the level of Multik. I am not sure if parallel processing should be turned on by default...

[Iaroslav Postovalov]

Which two lines? I don't get the context.

[altavir]

Check the new pull request. I've added a new version of LinearAlgebra for Jvm that introduces parallel processing for matrix building. It works quite well on dot operation.

[Iaroslav Postovalov]

So, it's just JVM's parallel stream? Then it's not surprising, because it's just MISD. Clearly should it been tried before...

[Iaroslav Postovalov]

We have spent many efforts to achieve SIMD with Viktor or ND4J.

[altavir]

Yes. I mean that usually, it is not effective for mathematical operations because the overhead from parallel processing is comparable to the benefit. But the dot operation specifically benefits a lot from it. Operations for each i and j could be done in parallel. And we do not need to make all operations parallel, only this one. We do it by creating a context for optimisation of a specific operation.

[altavir]

I checked the same context with other operations and it does not work well.

[Iaroslav Postovalov]

GPU computations are beneficial from parallelism exactly because of lower overhead you mentioned.

[altavir]

It is not lower. The computation itself is cheaper, but data transfer is much more expensive. GPU works well when you can load all your data at once and not update it.

[Iaroslav Postovalov]

Do you have a benchmark comparing kmath-multik, kmath-multik with parallel stream and pure multik?

[Iaroslav Postovalov]

Oh, I am wrong.

[Iaroslav Postovalov]

Clearly, multik has their own dot . And I'm sure kmath when wraps multik gives no or minimal overhead.

[altavir]

Pure multik is the same as KMath-multik since KMath is a thin wrapper on top. But yes. I've done exactly that. On Oracle GraalVM Multik is about 3-4 times faster than kmath with parallel processing. And about 20 times faster than Kmath without parallel processing.

[altavir]

At the same time KMath parallel is faster that both Ejml and TensorFlow-CPU

[Iaroslav Postovalov]

I think there's no use case for kmath-core with parallelism involved, when it's anyway worse than wrappers.

[altavir]

It is a good example. And it IS better than most of wrappers. Multik is good, but it is rather limited. Dot operations is practically the only non-trivial operation it could do.

[altavir]

So if one wants reasonable flexibility, but has a botleneck at dot operation, it could help. You just need to do a context switch.

[Iaroslav Postovalov]

If it's just an example, then more experiments should be done, with tensors, for instance.

[altavir]

Indeed. But tensors require a lot of clean up, which I can't do right now.