Documentation related to Dask, Distributed, and related packages. Utility functions commonly used by AICS projects.
- Distributed handler to manage various debugging or cluster configurations
- Documentation on example cluster deployments
Before we jump into quick starts there are some basic definitions to understand.
A task is a single static function to be processed. Simple enough. However, relevant to
AICS, is that when using aicsimageio
(and / or dask.array.Array
), your image (or
dask.array.Array
) is split up into many tasks. This is dependent on the image reader
and the size of the file you are reading. But in general it is safe to assume that each
image you read is split many thousands of tasks. If you want to see how many tasks your
image is split into you can either compute:
- Psuedo-code:
sum(2 * size(channel) for channel if channel not in ["Y", "X"])
- Dask graph length:
len(AICSImage.dask_data.__dask_graph__())
Apply a given function to the provided iterables as used as parameters to the function.
Given lambda x: x + 1
and [1, 2, 3]
, the result of map(func, *iterables)
in this
case would be [2, 3, 4]
. Usually, you are provided back an iterable of future
objects back from a map
operation. The results from the map operation are not
guaranteed to be in the order of the iterable that went in as operations are started as
resources become available and item to item variance may result in different output
ordering.
An object that will become available but is currently not defined. There is no guarantee
that the object is a valid result or an error and you should handle errors once the
future's state has resolved (usually this means after a gather
operation).
Block the process from moving forward until all futures are resolved. Control flow here would mean that you could potentially generate thousands of futures and keep moving on locally while those futures slowly resolve but if you ever want a hard stop and wait for some set of futures to complete, you would need gather them.
Dask tries to mirror the standard library concurrent.futures
wherever possible which
is what allows for this library to have simple wrappers around Dask to allow for easy
debugging as we are simply swapping out distributed.Client.map
with
concurrent.futures.ThreadPoolExecutor.map
for example. If at any point in your code
you don't want to use dask
for some reason or another, it is equally valid to use
concurrent.futures.ThreadPoolExecutor
or concurrent.futures.ProcessPoolExecutor
.
If you have an iterable (or iterables) that would result in less than hundreds of
thousands of tasks, it you can simply use the normal map
provided by the
DistributedHandler.client
.
Important Note: Notice, "... iterable that would result in less than hundreds
of thousands of tasks...". This is important because what happens when you try to map
over a thousand image paths, each which spawns an AICSImage
object. Each one adds
thousands more tasks to the scheduler to complete. This will break and you should look
to Large Iterable Batching instead.
from aics_dask_utils import DistributedHandler
# `None` address provided means use local machine threads
with DistributedHandler(None) as handler:
futures = handler.client.map(
lambda x: x + 1,
[1, 2, 3]
)
results = handler.gather(futures)
from distributed import LocalCluster
cluster = LocalCluster()
# Actual address provided means use the dask scheduler
with DistributedHandler(cluster.scheduler_address) as handler:
futures = handler.client.map(
lambda x: x + 1,
[1, 2, 3]
)
results = handler.gather(futures)
If you have an iterable (or iterables) that would result in more than hundreds of
thousands of tasks, you should use handler.batched_map
to reduce the load on the
client. This will batch your requests rather than send than all at once.
from aics_dask_utils import DistributedHandler
# `None` address provided means use local machine threads
with DistributedHandler(None) as handler:
results = handler.batched_map(
lambda x: x + 1,
range(1e9) # 1 billion
)
from distributed import LocalCluster
cluster = LocalCluster()
# Actual address provided means use the dask scheduler
with DistributedHandler(cluster.scheduler_address) as handler:
results = handler.batched_map(
lambda x: x + 1,
range(1e9) # 1 billion
)
Note: Notice that there is no handler.gather
call after batched_map
. This is
because batched_map
gathers results at the end of each batch rather than simply
returning their future's.
Stable Release: pip install aics_dask_utils
Development Head: pip install git+https://github.com/AllenCellModeling/aics_dask_utils.git
For full package documentation please visit AllenCellModeling.github.io/aics_dask_utils.
See CONTRIBUTING.md for information related to developing the code.
This README, provided tooling, and documentation are not meant to be all encompassing
of the various operations you can do with dask
and other similar computing systems.
For further reading go to dask.org.
Free software: Allen Institute Software License