High Performance Analytics Toolkit (HPAT) scales analytics/ML codes in Python to bare-metal cluster/cloud performance automatically. It compiles a subset of Python (Pandas/Numpy) to efficient parallel binaries with MPI, requiring only minimal code changes. HPAT is orders of magnitude faster than alternatives like Apache Spark.
HPAT's documentation can be found here.
HPAT can be installed in Anaconda environment easily (Linux/Mac/Windows):
conda create -n HPAT -c ehsantn -c anaconda -c conda-forge hpat
Windows installaton requires Intel MPI to be installed.
An HPAT docker image is also available for running containers. For example:
docker run -it ehsantn/hpat bash
Here is a Pi calculation example in HPAT:
import hpat
import numpy as np
import time
@hpat.jit
def calc_pi(n):
t1 = time.time()
x = 2 * np.random.ranf(n) - 1
y = 2 * np.random.ranf(n) - 1
pi = 4 * np.sum(x**2 + y**2 < 1) / n
print("Execution time:", time.time()-t1, "\nresult:", pi)
return pi
calc_pi(2 * 10**8)Save this in a file named pi.py and run (on 8 cores):
mpiexec -n 8 python pi.py
This should demonstrate about 100x speedup compared to regular Python version without @hpat.jit and mpiexec.
To use HPAT with Jupyter Notebook, install jupyter, ipython, and ipyparallel. Then, create a mpi profile for ipython:
ipython profile create --parallel --profile=mpi
Next, edit the ipcluster_config.py file. This file will be in your ipython directory in the profile_mpi directory. Your ipython directory is in your IPYTHONDIR environment variable if you have one defined and ~/.ipython if you don't have this variable defined. To the ipcluster_config.py file, add the following line:
c.IPClusterEngines.engine_launcher_class = 'MPIEngineSetLauncher'
Then, start the Jupyter notebook and click on IPython Clusters, select the number of engines (i.e., cores) you'd like to use and click Start next to the mpi profile. You can now run an HPAT function and the work will be distributed across the number of cores you selected on the current node.
If you wish to run across multiple nodes, you can add the following to ipcluster_config.py:
c.MPILauncher.mpi_args = ["-machinefile", "path_to_file/machinefile"]
This machinefile option is forwarded to mpi and the specified machine file in the second argument contains a list of machine names across which to distribute work. More information about the -machinefile option can be found here.
These academic papers describe the underlying methods in HPAT: