Scaling on Distributed Memory (Multiprocessing)
Note
Scikit-learn patching functionality in daal4py was deprecated and moved to a separate package, Intel(R) Extension for Scikit-learn*. All future patches will be available only in Intel(R) Extension for Scikit-learn*. Use the scikit-learn-intelex package instead of daal4py for the scikit-learn acceleration.
It’s Easy
daal4py operates in SPMD style (Single Program Multiple Data), which means your program is executed on several processes (e.g. similar to MPI). The use of MPI is not required for daal4py’s SPMD-mode to work, all necessary communication and synchronization happens under the hood of daal4py. It is possible to use daal4py and mpi4py in the same program, though.
Only very minimal changes are needed to your daal4py code to allow daal4py to run on a cluster of workstations. Initialize the distribution engine:
daalinit()
Add the distribution parameter to the algorithm construction:
kmi = kmeans_init(10, method="plusPlusDense", distributed=True)
When calling the actual computation each process expects an input file or input array/DataFrame. Your program needs to tell each process which file/array/DataFrame it should operate on. Like with other SPMD programs this is usually done conditionally on the process id/rank (‘daal4py.my_procid()’). Assume we have one file for each process, all having the same prefix ‘file’ and being suffixed by a number. The code could then look like this:
result = kmi.compute("file{}.csv", daal4py.my_procid())
The result of the computation will now be available on all processes.
Finally stop the distribution engine:
daalfini()
That’s all for the python code:
from daal4py import daalinit, daalfini, kmeans_init
daalinit()
kmi = kmeans_init(10, method="plusPlusDense", distributed=True)
result = kmi.compute("file{}.csv", daal4py.my_procid())
daalfini()
To actually get it executed on several processes use standard MPI mechanics, like:
mpirun -n 4 python ./kmeans.py
The binaries provided by Intel use the Intel® MPI library, but daal4py can also be compiled for any other MPI implementation.
Supported Algorithms and Examples
The following algorithms support distribution:
PCA (pca)
SVD (svd)
Linear Regression Training (linear_regression_training)
Ridge Regression Training (ridge_regression_training)
Multinomial Naive Bayes Training (multinomial_naive_bayes_training)
K-Means (kmeans_init and kmeans)
Correlation and Variance-Covariance Matrices (covariance)
Moments of Low Order (low_order_moments)
QR Decomposition (qr)