====== Performance studies ====== The objective of this study is to analyze xgboost's performance on the cluster. **Prerequisites** * Python Ananoda 3.7 * Xgboost (Installed with conda) * Mesoshared (CPU Xeon Platinum 8160, 2.1 GHZ) With 92 cores ===== How to===== Covert the notebook to python script: $ ipython nbconvert --to python 3.XGBoost ===== Results ===== ==== Multithreaded support ==== Xgboost is multithreaed. We control the number of threads by ''OMP_NUm_THREADS'' env variable For example to for 8 cores: $ module load lang/python/anaconda/3.7 $ OMP_NUM_TREADS=8 time ipython 3.XGBoost.py ^# Threads ^Time(m)^ |1|11m32s| |8|2m03s| |16|1m18s| |24|1m8s| |32|1m5s| |40|1m5s| |**48**|**1m5s**| |64|1m54s| |72|2m13s| |92|3m21s| **Conclusion** : Uper limit scalability is about **24 cores** ==== GPU support ==== We will use Volta GPU node. The package ''py-xgboost-gpu'' is installed in this node with ''conda'' $ qlogin -qvolta.q -l h_vmem=20G We need to modify our program to use GPU. I just added this param ''tree_method= 'gpu_hist''' in ''xgb_reg'' object. 150 xgb_reg = xgboost.XGBRegressor(learning_rate = 0.01, 151 max_depth = 10, 152 random_state=42, 153 n_estimators = 173, 154 tree_method= 'gpu_hist', 155 n_jobs=-1, 156 objective = 'count:poisson') We load anaconda $ export PATH=/opt/anaconda3/bin:$PATH Finaly we run our program: $ time ipython 3.XGBoostGPU.py **User time is about 38seconds** Using Multi GPU without sucess (Still reading docs :-) )