is multi-cpu supported by h2o-xgboost?

问题

Is there a configuration which allows to run H2OXGBoostEstimator in multithreading and not in the minimal config with one CPU, with h2o version 3.15.0.4035?

回答1:

xgboost implementation on H2O is multithreaded and like all other algorithms supported into H2O however it is platform dependent which is described into H2O documentation properly.

So if you try it on Linux, and have all supported libraries available then you will take advantage of distributed xgboost otherwise like OSX, you might get a single CPU fall back runtime. So it's all depend on which lib is loaded from your OS.

When H2O starts in the log you will see the following:

10-02 09:25:34.579 10.0.0.46:54321       54229  main      INFO: Registered 3 core extensions in: 57ms
10-02 09:25:34.580 10.0.0.46:54321       54229  main      INFO: Registered H2O core extensions: [Watchdog, XGBoost, KrbStandalone]
10-02 09:25:34.791 10.0.0.46:54321       54229  main      INFO: Registered: 161 REST APIs in: 211ms
10-02 09:25:34.791 10.0.0.46:54321       54229  main      INFO: Registered REST API extensions: [XGBoost, Algos, AutoML, Core V3, Core V4]

Then you will see if CPU/GPU is included as below:

10-02 09:23:49.952 10.0.0.46:54321       54143  FJ-1-5    INFO: No GPU (gpu_id: 0) found. Using CPU backend.

If you could run objdump or ldd command to see the libs loaded with H2O, you will have better idea what is missed which cause your xgboost runtime to be single CPU.

来源：https://stackoverflow.com/questions/46523998/is-multi-cpu-supported-by-h2o-xgboost