Making gridding faster

Hello everyone,

I would like to share with you my recent experience with gridding data and how using multiprocessing on a High-Performance Computing (HPC) cluster has dramatically reduced my processing time. In fact, I was able to grid 54 files of NEXRAD data in less than 10 minutes.

To achieve this, I used the multiprocessing library in Python, which allows multiple processes to run simultaneously, taking advantage of the cluster’s computing power. I want to emphasize that I did not try this approach on my local machine, and it is essential to have access to a powerful computing resource to benefit from multiprocessing.

import warnings
import pyart
import numpy as np
from matplotlib import pyplot as plt
import glob, os
from datetime import datetime
import time
import sys
import multiprocessing as mp

# Define a function that takes a file name and writes the GRID file
def process_file(nexrad_dir, rfile):
    print(f'Reading file: {rfile.split("/")[-1]}')
    radar =
    radar = filter_data(radar, "DBZ", "VEL")
    max_rng = 360.*1e3 # np.ceil(radar.range['data'].max())
    grid =,(61,721,721),
                       ((0.,15000.),(-max_rng, max_rng),(-max_rng, max_rng)), 
                                      fields=['REF', 'VEL', 'ZDR', 'RHO']
    print(f'Deleting: {rfile.split("/")[-1]}')
    del radar

    # Extract the NEXRAD file number from the file name
    file_number = rfile.split('NEXRAD')[-1]

    # Construct the full path to the GRID file using os.path.join()
    grid_file_path = os.path.join(nexrad_dir, f'GRID{file_number}')
    print(f'Saving Grid: {rfile.split("/")[-1]}')
    # Write the GRID file using, grid=grid)

def process_file_wrapper(args):
    return process_file(*args)

nexrad_dir = "/xxxxxxx/xxxxx/xxxxx/xxxx/obsdata/2022/NEXRAD/"
data_files = get_nexrad_data_files("20220330220000", "20220331040000")

# Create a pool of worker processes with 100 CPUs
pool = mp.Pool(processes=100)

# Map the data files to the process_file function to process them in parallel
results =, [(nexrad_dir, f) for f in data_files])

# Close the pool of worker processes
1 Like

Have you tried using dask and dask-jobqueue for your processing? The first makes mapping these radar processing tasks in parallel very easy with the use of a dask bag. The latter is a way to have it interact with slurm/PBS clusters from within Python so that you can map your task to multiple nodes. This is what we use all the time for our ARM processing.

For mapping PyDDA on a GPU cluster there is also dask-cuda, which will extend dask’s functionality to GPU-based nodes.

1 Like

Hi @rcjackson,

Thanks for suggesting dask and dask-jobqueue for processing radar data in parallel. I’ve had some errors in the past when trying to use them, but I’m willing to give it another shot. If you have any helpful resources or code to get me started, I’d appreciate it.

@syedhamidali - I recommend checking out the notebooks/talk I gave at one of the Dask Demo Days where I showcased how to use Py-ART/Xradar to scale analysis with Dask

In the readme, you will find the link to the youtube video, while the repository contains the notebooks and execution environment.

1 Like