Pandas df.iterrows() parallelization

Question

I would like to parallelize the following code:

for row in df.iterrows():
    idx = row[0]
    k = row[1][\'Chromosome\']
    start,end = row[1][\'Bin\'].spl         


        

Answer 1

As @Khris said in his comment, you should split up your dataframe into a few large chunks and iterate over each chunk in parallel. You could arbitrarily split the dataframe into randomly sized chunks, but it makes more sense to divide the dataframe into equally sized chunks based on the number of processes you plan on using. Luckily someone else has already figured out how to do that part for us:

# don't forget to import
import pandas as pd
import multiprocessing

# create as many processes as there are CPUs on your machine
num_processes = multiprocessing.cpu_count()

# calculate the chunk size as an integer
chunk_size = int(df.shape[0]/num_processes)

# this solution was reworked from the above link.
# will work even if the length of the dataframe is not evenly divisible by num_processes
chunks = [df.ix[df.index[i:i + chunk_size]] for i in range(0, df.shape[0], chunk_size)]

This creates a list that contains our dataframe in chunks. Now we need to pass it into our pool along with a function that will manipulate the data.

def func(d):
   # let's create a function that squares every value in the dataframe
   return d * d

# create our pool with `num_processes` processes
pool = multiprocessing.Pool(processes=num_processes)

# apply our function to each chunk in the list
result = pool.map(func, chunks)

At this point, result will be a list holding each chunk after it has been manipulated. In this case, all values have been squared. The issue now is that the original dataframe has not been modified, so we have to replace all of its existing values with the results from our pool.

for i in range(len(result)):
   # since result[i] is just a dataframe
   # we can reassign the original dataframe based on the index of each chunk
   df.ix[result[i].index] = result[i]

Now, my function to manipulate my dataframe is vectorized and would likely have been faster if I had simply applied it to the entirety of my dataframe instead of splitting into chunks. However, in your case, your function would iterate over each row of the each chunk and then return the chunk. This allows you to process num_process rows at a time.

def func(d):
   for row in d.iterrow():
      idx = row[0]
      k = row[1]['Chromosome']
      start,end = row[1]['Bin'].split('-')

      sequence = sequence_from_coordinates(k,1,start,end) #slow download form http
      d.set_value(idx,'GC%',gc_content(sequence,percent=False,verbose=False))
      d.set_value(idx,'G4 repeats', sum([len(list(i)) for i in g4_scanner(sequence)]))
      d.set_value(idx,'max flexibility',max([item[1] for item in dna_flex(sequence,verbose=False)]))
   # return the chunk!
   return d

Then you reassign the values in the original dataframe, and you have successfully parallelized this process.

How Many Processes Should I Use?

Your optimal performance is going to depend on the answer to this question. While "ALL OF THE PROCESSES!!!!" is one answer, a better answer is much more nuanced. After a certain point, throwing more processes at a problem actually creates more overhead than it's worth. This is known as Amdahl's Law. Again, we are fortunate that others have already tackled this question for us:

1. Python multiprocessing's Pool process limit
2. How many processes should I run in parallel?

A good default is to use multiprocessing.cpu_count(), which is the default behavior of multiprocessing.Pool. According to the documentation "If processes is None then the number returned by cpu_count() is used." That's why I set num_processes at the beginning to multiprocessing.cpu_count(). This way, if you move to a beefier machine, you get the benefits from it without having to change the num_processes variable directly.

Answer 2

A faster way (about 10% in my case):

Main differences to accepted answer: use pd.concat and np.array_split to split and join the dataframre.

import multiprocessing
import numpy as np


def parallelize_dataframe(df, func):
    num_cores = multiprocessing.cpu_count()-1  #leave one free to not freeze machine
    num_partitions = num_cores #number of partitions to split dataframe
    df_split = np.array_split(df, num_partitions)
    pool = multiprocessing.Pool(num_cores)
    df = pd.concat(pool.map(func, df_split))
    pool.close()
    pool.join()
    return df

where func is the function you want to apply to df. Use partial(func, arg=arg_val) for more that one argument.

Answer 3

Consider using dask.dataframe, as e.g. shown in this example for a similar question: https://stackoverflow.com/a/53923034/4340584

import dask.dataframe as ddf
df_dask = ddf.from_pandas(df, npartitions=4)   # where the number of partitions is the number of cores you want to use
df_dask['output'] = df_dask.apply(lambda x: your_function(x), meta=('str')).compute(scheduler='multiprocessing')