The Java binding calls the same fastLowess Rust core as every other binding (C++, Python, R, Julia, Node.js, WASM, Go) through a thin JNI layer, so its raw fit performance is governed by the same underlying engine — see the C++ binding’s benchmarks for representative numbers (R baseline vs. serial vs. parallel, across scale/fraction/iteration/pathological-data scenarios).
A dedicated JMH benchmark harness (wired into benchmarks/compare.py alongside the R/Rust comparisons) has not been added yet — this is a good follow-up once the binding has seen some real-world use. In the meantime, JUnit tests combined with manual timing (or JMH) can be used for local relative comparisons between option configurations (e.g. serial vs. parallel(true), different fraction/iterations values).