Is there a preferred jdk, garbage collector, etc. ...
# mathematicsa
alex cole
01/14/2021, 12:05 AMIs there a preferred jdk, garbage collector, etc. that is best used for numerical loads?
a
altavir
01/14/2021, 6:30 AMGC is not a problem usually. In my experience GraalVM gives the best results so far.
i
Iaroslav Postovalov
01/14/2021, 6:55 AMThe point about Graal is correct. As for collector, choosing of it depends not on type of computations, but on expected latency, size of data set, number of CPU cores etc.
Iaroslav Postovalov
01/14/2021, 6:59 AMIf you tell some more information about your workload, I will give my opinion
a
alex cole
01/14/2021, 3:49 PMThis would be N-Body problems with vector fields. Starting off with a Euler gravitational simulation, still trying to figure out particle mesh. I haven’t used graal before, and I like to profile on my Pi 4 and I’ll run the full simulation on an AWS server.
a
altavir
01/14/2021, 3:52 PMI do not think that GC will be an issue. GraalVM is in general better for boxing avoidance, but to my surprise it is significantly better, that OpenJ9 in a lot of other tasks.
a
alex cole
01/14/2021, 3:56 PMDo you have a benchmark that is good at comparing different set ups between each jvm/hotspot/ garbage collectors to test against graal? Also would you happen to have a list of test loads for benchmarking numerical libraries?
a
altavir
01/14/2021, 4:00 PMThere are some benchmarks in this thread: https://discuss.kotlinlang.org/t/using-lambda-and-method-reference-differ-significantly-in-performance-with-java-8-streams/19306/11. I actually did some tests for kmath and GraalVM gave the best results, but I am not sure if I saved it somewhere. Also there are a lot of benchmarks in ojalgo blog: https://github.com/optimatika/ojAlgo. But Ojalglo is super-optimized by itself, so VM does not play a huge role there.
a
alex cole
01/14/2021, 4:10 PMI may actually have to use that library or atleast wrap it.
a
altavir
01/14/2021, 4:52 PMThere is the ongoing work on a wrapper https://github.com/mipt-npm/kmath/tree/ojalgo by @Iaroslav Postovalov, if you need it, you can cooperate and contribute. But I do not think that you will get a significant perforamce boost for n-body compared to other libraries like commons-math .
a
alex cole
01/14/2021, 4:58 PMI wrote a library just to make the vector math easily and have a familiar feel from matlab and numpy. Primarily I just wanted to directly operate on vectors to reduce math mistakes and improve upon an numerical methods library. https://github.com/jalexcole/numeric
a
altavir
01/15/2021, 6:17 AMI do not think that numpy approach is good for kotlin. Anyway, there is koma and now MultiK for that.
a
alex cole
01/15/2021, 11:48 PMUnder the hood, why would a numpy approach lead to poor performance?
a
altavir
01/16/2021, 5:17 AMI never said anything about performance, but performance is also an issue. We had a discussion about that yesterday with @Ролан. Suppose you want to compute an integral of a function. You can't use the actual lambda in numpy since lambda is really slow and you need for all operations to be don via numpy arrays. But numpy arrays are also slow on large expressions because you need to create an array for each operation instead of inlining the function.
In general, it is possible to do very fast calculus in a numpy-style, if you know which things to avoid, but my main concern is not the speed, but extensibility and maintainability.
a
alex cole
01/16/2021, 4:00 PMI have noticed the maintainability part by the polynomial Class and being able to add types to my ArrayND class for ComplexDoubles.
4 Views