CuML -- sklearn -- sklearnex NearestNeighbors Benchmarks

benchmark.py

import sys                                                                                                                                                                                                                                                    
import cudf                                                                                                                                                                                                                                                   
import joblib                                                                                                                                                                                                                                                 
import numpy as np                                                                                                                                                                                                                                            
import time
import sklearn

if "--sklearnex" in sys.argv:
    import sklearnex
    sklearnex.patch_sklearn()

sklearn.show_versions()

from cuml.neighbors import NearestNeighbors as CUNN
from sklearn.neighbors import NearestNeighbors as SKNN
from sklearn.datasets import make_blobs


params = dict(n_neighbors=1024, n_jobs=-1, algorithm='brute')

m = joblib.Memory(location='/tmp/jjerphan')

make_blobs = m.cache(make_blobs)

X, _ = make_blobs(
    n_samples=65_000,
    centers=5,
    n_features=256,
    random_state=42,
)

Y, _ = make_blobs(
    n_samples=65_000,
    centers=5,
    n_features=256,
    random_state=43,
)


if "--float32":
    X = X.astype(np.float32)
    Y = Y.astype(np.float32)

X_cudf = cudf.DataFrame(X.astype(np.float32))
Y_cudf = cudf.DataFrame(Y.astype(np.float32))

cu_durations = []
sk_durations = []

cu_model = CUNN(**params)

for i in range(5):
    start = time.perf_counter()
    cu_model.fit(X_cudf)
    cu_model.kneighbors(Y_cudf)
    end = time.perf_counter()

    cu_duration = end - start
    cu_durations.append(cu_duration)
    print("CUDA:", cu_duration)

model = SKNN(**params)

for i in range(3):
    start = time.perf_counter()
    model.fit(X)
    model.kneighbors(Y)
    end = time.perf_counter()

    sk_duration = end - start
    sk_durations.append(sk_duration)

    print("skl:", sk_duration)


print("Ratio:",  np.min(sk_duration) / np.min(cu_duration))

(rapids-21.10+intelex) ⋅[email protected] ~ 🌍
$ srun --partition=gpu --gpus=1 -c 64 `where python` bench_knn.py --sklearnex --float32                                                                                                                                                                      
srun: job 300314 queued and waiting for resources                                                                                                                                                                                                             
srun: job 300314 has been allocated resources                                                                                                                                                                                                                 
Intel(R) Extension for Scikit-learn* enabled (https://github.com/intel/scikit-learn-intelex)                                                                                                                                                                  
                                                                                                                                                                                                                                                              
System:                                                                                                                                                                                                                                                       
    python: 3.8.12 | packaged by conda-forge | (default, Oct 12 2021, 21:59:51)  [GCC 9.4.0]                                                                                                                                                                  
executable: /home/parietal/jjerphan/.local/miniconda3/envs/rapids-21.10+intelex/bin/python                                                                                                                                                                    
   machine: Linux-4.18.0-240.1.1.el8_3.x86_64-x86_64-with-glibc2.10                                                                                                                                                                                           
                                                                                                                                                                                                                                                              
Python dependencies:                                                                                                                                                                                                                                          
          pip: 21.3                                                                                                                                                                                                                                           
   setuptools: 58.2.0                                                                                                                                                                                                                                         
      sklearn: 0.24.2                                                                                                                                                                                                                                         
        numpy: 1.21.2                                                                                                                                                                                                                                         
        scipy: 1.7.1                                                                                                                                                                                                                                          
       Cython: None                                                                                                                                                                                                                                           
       pandas: 1.3.3                                                                                                                                                                                                                                          
   matplotlib: None                                                                                                                                                                                                                                           
       joblib: 1.1.0                                                                                                                                                                                                                                          
threadpoolctl: 2.1.0                                                                                                                                                                                                                                          
                                                                                                                                                                                                                                                              
Built with OpenMP: True                                                                                                                                                                                                                                       
CUDA: 1.3312459930311888                                                                                                                                                                                                                                      
CUDA: 0.7997059129411355                                                                                                                                                                                                                                      
CUDA: 0.8033369790064171                                                                                                                                                                                                                                      
CUDA: 0.9464123289799318                                                                                                                                                                                                                                      
CUDA: 0.8314744649687782                                                                                                                                                                                                                                      
skl: 2.346788124064915                                                                                                                                                                                                                                        
skl: 2.2880533149000257                                                                                                                                                                                                                                       
skl: 2.2813473369460553                                                                                                                                                                                                                                       
Ratio: 2.7437370996495

(rapids-21.10+intelex) ⋅[email protected] ~ 🌍
$ srun --partition=gpu --gpus=1 -c 64 `where python` bench_knn.py --sklearnex          
srun: job 300370 queued and waiting for resources
srun: job 300370 has been allocated resources
Intel(R) Extension for Scikit-learn* enabled (https://github.com/intel/scikit-learn-intelex)

System:
    python: 3.8.12 | packaged by conda-forge | (default, Oct 12 2021, 21:59:51)  [GCC 9.4.0]
executable: /home/parietal/jjerphan/.local/miniconda3/envs/rapids-21.10+intelex/bin/python
   machine: Linux-4.18.0-240.1.1.el8_3.x86_64-x86_64-with-glibc2.10

Python dependencies:
          pip: 21.3
   setuptools: 58.2.0
      sklearn: 0.24.2
        numpy: 1.21.2
        scipy: 1.7.1
       Cython: None
       pandas: 1.3.3
   matplotlib: None
       joblib: 1.1.0
threadpoolctl: 2.1.0

Built with OpenMP: True
CUDA: 1.3430722079938278
CUDA: 0.7992004080442712
CUDA: 0.801696153008379
CUDA: 0.9491368799936026
CUDA: 0.8242675189394504
skl: 2.288681124104187
skl: 2.2949441720265895
skl: 2.27677305706311
Ratio: 2.7621773329034442