Machine learning algorithm with 20 threads. Machine is like: "bitch please" (and runs the algo clean until the end)

Increase parameter to 40 threads. NullPointerExceptions. NullPointerExceptions everywhere.

How many cores does the machine it's running on have? 40 threads seems a bit much, you may be better off using a hand full of threads in a thread pool layout and feed it with a work queue.

@hexc is right.

Large number of threads does not improve performance. If you exceed the number of cores, you will degrade performance.

I have 8 cpu cores. It did manage to run with 40 threads for specific and lighter datasets. I think there are some methods I have synchronize. Or maybe I changed aomwthing in the code by accident in between these runs.

That could be but remember, if you have more threads than cores you usually get worse performance. The one exception being a Hyperthreaded processor. With one of those, you usually can get away with a bit more but I'd recommend benchmarking the code to find the sweet spot for the workload.