Returning a generator in joblib.Parallel

This example illustrates memory optimization enabled by using joblib.Parallel to get a generator on the outputs of parallel jobs. We first create tasks that return results with large memory footprints. If we call Parallel for several of these tasks directly, we observe a high memory usage, as all the results are held in RAM before being processed

Using return_as='generator' allows to progressively consume the outputs as they arrive and keeps the memory at an acceptable level.

In this case, the output of the Parallel call is a generator that yields the results in the order the tasks have been submitted with. If the order of the tasks does not matter (for instance if they are consumed by a commutative aggregation function), then using return_as='generator_unordered' can be even more efficient.

MemoryMonitor helper

The following class is an helper to monitor the memory of the process and its children in another thread, so we can display it afterward.

We will use psutil to monitor the memory usage in the code. Make sure it is installed with pip install psutil for this example.

import time
from psutil import Process
from threading import Thread


class MemoryMonitor(Thread):
    """Monitor the memory usage in MB in a separate thread.

    Note that this class is good enough to highlight the memory profile of
    Parallel in this example, but is not a general purpose profiler fit for
    all cases.
    """
    def __init__(self):
        super().__init__()
        self.stop = False
        self.memory_buffer = []
        self.start()

    def get_memory(self):
        "Get memory of a process and its children."
        p = Process()
        memory = p.memory_info().rss
        for c in p.children():
            memory += c.memory_info().rss
        return memory

    def run(self):
        memory_start = self.get_memory()
        while not self.stop:
            self.memory_buffer.append(self.get_memory() - memory_start)
            time.sleep(0.2)

    def join(self):
        self.stop = True
        super().join()

Save memory by consuming the outputs of the tasks as fast as possible

We create a task whose output takes about 15MB of RAM.

import numpy as np


def return_big_object(i):
    time.sleep(.1)
    return i * np.ones((10000, 200), dtype=np.float64)

We create a reduce step. The input will be a generator on big objects generated in parallel by several instances of return_big_object.

def accumulator_sum(generator):
    result = 0
    for value in generator:
        result += value
        print(".", end="", flush=True)
    print("")
    return result

We process many of the tasks in parallel. If return_as="list" (default), we should expect a usage of more than 2GB in RAM. Indeed, all the results are computed and stored in res before being processed by accumulator_sum and collected by the gc.

from joblib import Parallel, delayed

monitor = MemoryMonitor()
print("Running tasks with return_as='list'...")
res = Parallel(n_jobs=2, return_as="list")(
    delayed(return_big_object)(i) for i in range(150)
)
print("Accumulate results:", end='')
res = accumulator_sum(res)
print('All tasks completed and reduced successfully.')

# Report memory usage
del res  # we clean the result to avoid memory border effects
monitor.join()
peak = max(monitor.memory_buffer) / 1e9
print(f"Peak memory usage: {peak:.2f}GB")

Running tasks with return_as='list'...
Accumulate results:......................................................................................................................................................
All tasks completed and reduced successfully.
Peak memory usage: 2.49GB

If we use return_as="generator", res is simply a generator on the results that are ready. Here we consume the results as soon as they arrive with the accumulator_sum and once they have been used, they are collected by the gc. The memory footprint is thus reduced, typically around 300MB.

monitor_gen = MemoryMonitor()
print("Create result generator with return_as='generator'...")
res = Parallel(n_jobs=2, return_as="generator")(
    delayed(return_big_object)(i) for i in range(150)
)
print("Accumulate results:", end='')
res = accumulator_sum(res)
print('All tasks completed and reduced successfully.')

# Report memory usage
del res  # we clean the result to avoid memory border effects
monitor_gen.join()
peak = max(monitor_gen.memory_buffer) / 1e6
print(f"Peak memory usage: {peak:.2f}MB")

Create result generator with return_as='generator'...
Accumulate results:......................................................................................................................................................
All tasks completed and reduced successfully.
Peak memory usage: 217.11MB

We can then report the memory usage accross time of the two runs using the MemoryMonitor.

In the first case, as the results accumulate in res, the memory grows linearly and it is freed once the accumulator_sum function finishes.

In the second case, the results are processed by the accumulator as soon as they arrive, and the memory does not need to be able to contain all the results.

import matplotlib.pyplot as plt
plt.figure(0)
plt.semilogy(
    np.maximum.accumulate(monitor.memory_buffer),
    label='return_as="list"'
)
plt.semilogy(
    np.maximum.accumulate(monitor_gen.memory_buffer),
    label='return_as="generator"'
)
plt.xlabel("Time")
plt.xticks([], [])
plt.ylabel("Memory usage")
plt.yticks([1e7, 1e8, 1e9], ['10MB', '100MB', '1GB'])
plt.legend()
plt.show()

It is important to note that with return_as="generator", the results are still accumulated in RAM after computation. But as we asynchronously process them, they can be freed sooner. However, if the generator is not consumed the memory still grows linearly.

Further memory efficiency for commutative aggregation

There is still room for improving the relief on memory allocation we get using return_as="generator". Indeed, notice how the generator of the previous example respects the order the tasks have been submitted with. This behavior can cause a build up in memory of results waiting to be consumed, in case some tasks finished before other tasks despite being submitted later. The corresponding results will be kept in memory until the slower tasks submitted earlier are done and have been iterated over.

In case the downstream consumer of the results is reliant on the assumption that the results are yielded in the same order that the tasks were submitted, it can’t be helped. But in our example, since the + operator is commutative, the function accumulator_sum does not need the generator to return the results with any particular order. In this case it’s safe to use the option return_as="generator_unordered", so that the results are returned as soon as a task is completed, ignoring the order of task submission.

Beware that the downstream consumer of the results must not expect them be returned with any deterministic or predictable order at all, since the progress of the tasks can depend on the availability of the workers, which can be affected by external events, such as system load, implementation details in the backend, etc.

To better highlight improvements in memory usage when using the parameter return_as="generator_unordered", let’s explcitly add delay in some of the submitted tasks.

def return_big_object_delayed(i):
    if (i + 20) % 60:
        time.sleep(0.1)
    else:
        time.sleep(5)
    return i * np.ones((10000, 200), dtype=np.float64)

Let’s check memory usage when using return_as="generator"…

monitor_delayed_gen = MemoryMonitor()
print("Create result generator on delayed tasks with return_as='generator'...")
res = Parallel(n_jobs=2, return_as="generator")(
    delayed(return_big_object_delayed)(i) for i in range(150)
)
print("Accumulate results:", end='')
res = accumulator_sum(res)
print('All tasks completed and reduced successfully.')

# Report memory usage
del res  # we clean the result to avoid memory border effects
monitor_delayed_gen.join()
peak = max(monitor_delayed_gen.memory_buffer) / 1e6
print(f"Peak memory usage: {peak:.2f}MB")

Create result generator on delayed tasks with return_as='generator'...
Accumulate results:......................................................................................................................................................
All tasks completed and reduced successfully.
Peak memory usage: 623.90MB

If we use return_as="generator_unordered", res will not enforce any order when returning the results, and will simply enable iterating on the results as soon as it’s available. The peak memory usage is now controlled to an even lower level, since that results can be consumed immediately rather than being delayed by the compute of slower tasks that have been submitted earlier.

monitor_delayed_gen_unordered = MemoryMonitor()
print(
  "Create result generator on delayed tasks with "
  "return_as='generator_unordered'..."
)
res = Parallel(n_jobs=2, return_as="generator_unordered")(
    delayed(return_big_object_delayed)(i) for i in range(150)
)
print("Accumulate results:", end='')
res = accumulator_sum(res)
print('All tasks completed and reduced successfully.')

# Report memory usage
del res  # we clean the result to avoid memory border effects
monitor_delayed_gen_unordered.join()
peak = max(monitor_delayed_gen_unordered.memory_buffer) / 1e6
print(f"Peak memory usage: {peak:.2f}MB")

Create result generator on delayed tasks with return_as='generator_unordered'...
Accumulate results:......................................................................................................................................................
All tasks completed and reduced successfully.
Peak memory usage: 319.08MB

Notice how the plot for 'return_as="generator' now shows a high memory usage plateau when slow jobs cause a congestion of intermediate results waiting in RAM before in-order aggregation. This high memory usage is never observed when using 'return_as="generator_unordered".

plt.figure(1)
plt.semilogy(
    np.maximum.accumulate(monitor_delayed_gen.memory_buffer),
    label='return_as="generator"'
)
plt.semilogy(
    np.maximum.accumulate(monitor_delayed_gen_unordered.memory_buffer),
    label='return_as="generator_unordered"'
)
plt.xlabel("Time")
plt.xticks([], [])
plt.ylabel("Memory usage")
plt.yticks([1e7, 1e8, 1e9], ['10MB', '100MB', '1GB'])
plt.legend()
plt.show()