[jvm-package] xgboost4j native method create many threads but not closed

We have a java service using xgboost model to predict. The xgboost model is trained by algorithm engineers with Python and then we use xgboost4j 0.72 to load model and predict online.
After the service starts for a while, we found too many threads (3000+) in system by `ps -eLf` while`jstack pid` shows just 300+ threads.
We suspected and found that the JNI call (C++ code) to predict created `nthread` threads nearly every time.
Is there anybody who met the same problem and can help us?
thanks.