Introduction to Dask

Parallel Programming with Dask in Python

James Fulton

Climate Informatics Researcher

Speeding up computations using multiple cores

Computers have multiple cores
Code needs to be written to use them
The Dask package can be used to do this
Complete our computations faster

Concurrent programming

A diagram which shows a list of tasks and one path going through them.

Multi-threading

A list of tasks which is split up into two.

Multi-threading

Two sets of tasks being sent to two different CPU cores.

Multi-threading

The two sets of tasks are run from within the same Python process.

Parallel processing

The two sets of tasks are now run from two different Python processes.

Parallel programming

Multi-threading

Two sets of tasks run by two CPU cores which are run inside the same Python process.

Parallel processing

Two sets of tasks run by two CPU cores which are run inside two different Python processes.

Lazy evaluation

Computations are not run until the moment the result is needed
The steps required to compute the result are stored for later
Dask splits the tasks between threads or processes

Dask delayed

from dask import delayed

def my_square_function(x):
    return x**2


# Create delayed version of above function
delayed_square_function = delayed(my_square_function)

Dask delayed

from dask import delayed

def my_square_function(x):
    return x**2

# Create delayed version of above function
delayed_square_function = delayed(my_square_function)

# Use the delayed function with input 4
delayed_result = delayed_square_function(4)


# Print the delayed answer
print(delayed_result)

Delayed('my_square_function-7f71b132-70a9-457a-aa52-604e8c34f8a7')

Dask delayed

from dask import delayed

def my_square_function(x):
    return x**2

# Delay and use function
delayed_result = delayed(my_square_function)(4)

print(delayed_result)

Delayed('my_square_function-7f71b132-70a9-457a-aa52-604e8c34f8a7')

Computing the answer

from dask import delayed

def my_square_function(x):
    return x**2

delayed_result = delayed(my_square_function)(4)

real_result = delayed_result.compute() # <- This line is where the calculation happens

# Print the answer
print(real_result)

Using operations on delayed objects

delayed_result1 = delayed(my_square_function)(4)

# Math operations return delayed object
delayed_result2 = (4 + delayed_result1) * 5

print(delayed_result2.compute())

Lazy evaluation

x_list = [30, 85, 14, 12, 27, 62, 89, 15, 78,  0]

sum_of_squares = 0

for x in x_list:
    # Square and add numbers
    sum_of_squares += delayed(my_square_function)(x)

Lazy evaluation

x_list = [30, 85, 14, 12, 27, 62, 89, 15, 78,  0]

sum_of_squares = 0

for x in x_list:
    # Square and add numbers
    sum_of_squares += delayed(my_square_function)(x)

result = sum_of_squares.compute()

# Print the answer
print(result)

Sharing computation

delayed_intermediate = delayed(my_square_function)(3)

# These two results both use delayed_intermediate
delayed_result1 = delayed_intermediate - 5
delayed_result2 = delayed_intermediate + 4

# delayed_3_squared will be computed twice
print('delayed_result1:', delayed_result1.compute())
print('delayed_result2:', delayed_result2.compute())

delayed_result1: 4
delayed_result2: 13

Sharing computation

import dask

# delayed_intermediate will be computed once
comp_result1, comp_result2 = dask.compute(delayed_result1, delayed_result2)

print('comp_result1:', comp_result1)
print('comp_result2:', comp_result2)

delayed_result1: 4
delayed_result2: 13

Let's practice!

Parallel Programming with Dask in Python