Diagnostics

Profiling parallel code can be tricky, but dask.diagnostics provides functionality to aid in profiling and inspecting dask graph execution.

Scheduler Callbacks

Schedulers based on dask.async.get_async (currently dask.async.get_sync, dask.threaded.get, and dask.multiprocessing.get) accept five callbacks, allowing for inspection of scheduler execution.

The callbacks are:

  1. start(dsk)

    Run at the beginning of execution, right before the state is initialized. Receives the dask graph.

  2. start_state(dsk, state)

    Run at the beginning of execution, right after the state is initialized. Receives the dask graph and scheduler state.

  3. pretask(key, dsk, state)

    Run every time a new task is started. Receives the key of the task to be run, the dask graph, and the scheduler state.

  4. posttask(key, result, dsk, state, id)

    Run every time a task is finished. Receives the key of the task that just completed, the result, the dask graph, the scheduler state, and the id of the worker that ran the task.

  5. finish(dsk, state, errored)

    Run at the end of execution, right before the result is returned. Receives the dask graph, the scheduler state, and a boolean indicating whether or not the exit was due to an error.

These are internally represented as tuples of length 5, stored in the order presented above. Callbacks for common use cases are provided in dask.diagnostics.

Progress Bar

The ProgressBar class builds on the scheduler callbacks described above to display a progress bar in the terminal or notebook during computation. This can be a nice feedback during long running graph execution. It can be used as a context manager around calls to get or compute to profile the computation:

>>> from dask.diagnostics import ProgressBar
>>> a = da.random.normal(size=(10000, 10000), chunks=(1000, 1000))
>>> res = a.dot(a.T).mean(axis=0)

>>> with ProgressBar():
...     out = res.compute()
[########################################] | 100% Completed | 17.1 s

Or registered globally using the register method.

>>> pbar = ProgressBar()
>>> pbar.register()
>>> out = res.compute()
[########################################] | 100% Completed | 17.1 s

To unregister from the global callbacks, call the unregister method:

>>> pbar.unregister()

Profiling

Dask provides a few tools for profiling execution. As with the ProgressBar, they each can be used as context managers, or registered globally.

Profiler

The Profiler class is used to profile dask execution at the task level. During execution it records the following information for each task:

  1. Key
  2. Task
  3. Start time in seconds since the epoch
  4. Finish time in seconds since the epoch
  5. Worker id

ResourceProfiler

The ResourceProfiler class is used to profile dask execution at the resource level. During execution it records the following information for each timestep

  1. Time in seconds since the epoch
  2. Memory usage in MB
  3. % CPU usage

The default timestep is 1 second, but can be set manually using the dt keyword.

>>> from dask.diagnostics import ResourceProfiler
>>> rprof = ResourceProfiler(dt=0.5)

CacheProfiler

The CacheProfiler class is used to profile dask execution at the scheduler cache level. During execution it records the following information for each task:

  1. Key
  2. Task
  3. Size metric
  4. Cache entry time in seconds since the epoch
  5. Cache exit time in seconds since the epoch

Where the size metric is the output of a function called on the result of each task. The default metric is to count each task (metric is 1 for all tasks). Other functions may be used as a metric instead through the metric keyword. For example, the nbytes function found in cachey can be used to measure the number of bytes in the scheduler cache:

>>> from dask.diagnostics import CacheProfiler
>>> from cachey import nbytes
>>> cprof = CacheProfiler(metric=nbytes)

Example

As an example to demonstrate using the diagnostics, we’ll profile some linear algebra done with dask.array. We’ll create a random array, take its QR decomposition, and then reconstruct the initial array by multiplying the Q and R components together. Note that since the profilers (and all diagnostics) are just context managers, multiple profilers can be used in a with block:

>>> import dask.array as da
>>> from dask.diagnostics import Profiler, ResourceProfiler, CacheProfiler
>>> a = da.random.random(size=(10000, 1000), chunks=(1000, 1000))
>>> q, r = da.linalg.qr(a)
>>> a2 = q.dot(r)

>>> with Profiler() as prof, ResourceProfiler(dt=0.25) as rprof,
...         CacheProfiler() as cprof:
...     out = a2.compute()

The results of each profiler are stored in their results attribute as a list of namedtuple objects:

>>> prof.results[0]
TaskData(key=('tsqr-8d16e396b237bf7a731333130d310cb9_QR_st1', 5, 0),
         task=(qr, (_apply_random, 'random_sample', 1060164455, (1000, 1000), (), {})),
         start_time=1454368444.493292,
         end_time=1454368444.902987,
         worker_id=4466937856)

>>> rprof.results[0]
ResourceData(time=1454368444.078748, mem=74.100736, cpu=0.0)

>>> cprof.results[0]
CacheData(key=('tsqr-8d16e396b237bf7a731333130d310cb9_QR_st1', 7, 0),
          task=(qr, (_apply_random, 'random_sample', 1310656009, (1000, 1000), (), {})),
          metric=1,
          cache_time=1454368444.49662,
          free_time=1454368446.769452)

These can be analyzed separately, or viewed in a bokeh plot using the provided visualize method on each profiler:

>>> prof.visualize()

To view multiple profilers at the same time, the dask.diagnostics.visualize function can be used. This takes a list of profilers, and creates a vertical stack of plots aligned along the x-axis:

>>> from dask.diagnostics import visualize
>>> visualize([prof, rprof, cprof])

Looking at the above figure, from top to bottom:

  1. The results from the Profiler object. This shows the execution time for each task as a rectangle, organized along the y-axis by worker (in this case threads). Similar tasks are grouped by color, and by hovering over each task one can see the key and task that each block represents.
  2. The results from the ResourceProfiler object. This shows two lines, one for total CPU percentage used by all the workers, and one for total memory usage.
  3. The results from the CacheProfiler object. This shows a line for each task group, plotting the sum of the current metric in the cache against time. In this case it’s the default metric (count), and the lines represent the number of each object in the cache at time. Note that the grouping and coloring is the same as for the Profiler plot, and that the task represented by each line can be found by hovering over the line.

From these plots we can see that the initial tasks (calls to numpy.random.random and numpy.linalg.qr for each chunk) are run concurrently, but only use slightly more than 100% CPU. This is because the call to numpy.linalg.qr currently doesn’t release the global interpreter lock, so those calls can’t truly be done in parallel. Next, there’s a reduction step where all the blocks are combined. This requires all the results from the first step to be held in memory, as shown by the increased number of results in the cache, and increase in memory usage. Immediately after this task ends, the number of elements in the cache decreases, showing that they were only needed for this step. Finally, there’s an interleaved set of calls to dot and sum. Looking at the CPU plot shows that these run both concurrently and in parallel, as the CPU percentage spikes up to around 350%.

Custom Callbacks

Custom diagnostics can be created using the callback mechanism described above. To add your own, subclass the Callback class, and define your own methods. Here we create a class that prints the name of every key as it’s computed:

from dask.callbacks import Callback
class PrintKeys(Callback):
    def _pretask(self, key, dask, state):
        """Print the key of every task as it's started"""
        print("Computing: {0}!".format(repr(key)))

This can now be used as a context manager during computation:

>>> from operator import add, mul
>>> dsk = {'a': (add, 1, 2), 'b': (add, 3, 'a'), 'c': (mul, 'a', 'b')}
>>> with PrintKeys():
...     get(dsk, 'c')
Computing 'a'!
Computing 'b'!
Computing 'c'!

Alternatively, functions may be passed in as keyword arguments to Callback:

>>> def printkeys(key, dask, state):
...    print("Computing: {0}!".format(repr(key)))
>>> with Callback(pretask=printkeys):
...     get(dsk, 'c')
Computing 'a'!
Computing 'b'!
Computing 'c'!