This document describes the policies used to select the preference of tasks and to select the preference of workers used by Dask’s distributed scheduler. For more information on how this these policies are enacted efficiently see Scheduling State.
When a task transitions from waiting to a processing state we decide a suitable worker for that task. If the task has significant data dependencies or if the workers are under heavy load then this choice of worker can strongly impact global performance. Currently workers for tasks are determined as follows:
If the task has no major dependencies and no restrictions then we find the least occupied worker.
Otherwise, if a task has user-provided restrictions (for example it must run on a machine with a GPU) then we restrict the available pool of workers to just that set, otherwise we consider all workers
From among this pool of workers we determine the workers to whom the least amount of data would need to be transferred.
We break ties by choosing the worker that currently has the fewest tasks, counting both those tasks in memory and those tasks processing currently.
This process is easy to change (and indeed this document may be outdated). We
encourage readers to inspect the
decide_worker function in scheduler.py
Decide which worker should take task ts.
We often have a choice between running many valid tasks. There are a few competing interests that might motivate our choice:
Run tasks on a first-come-first-served basis for fairness between multiple clients
Run tasks that are part of the critical path in an effort to reduce total running time and minimize straggler workloads
Run tasks that allow us to release many dependencies in an effort to keep the memory footprint small
Run tasks that are related so that large chunks of work can be completely eliminated before running new chunks of work
Accomplishing all of these objectives simultaneously is impossible. Optimizing for any of these objectives perfectly can result in costly overhead. The heuristics with the scheduler do a decent but imperfect job of optimizing for all of these (they all come up in important workloads) quickly.
Last in, first out¶
When a worker finishes a task the immediate dependencies of that task get top priority. This encourages a behavior of finishing ongoing work immediately before starting new work. This often conflicts with the first-come-first-served objective but often results in shorter total runtimes and significantly reduced memory footprints.
Break ties with children and depth¶
Often a task has multiple dependencies and we need to break ties between them with some other objective. Breaking these ties has a surprisingly strong impact on performance and memory footprint.
When a client submits a graph we perform a few linear scans over the graph to determine something like the number of descendants of each node (not quite, because it’s a DAG rather than a tree, but this is a close proxy). This number can be used to break ties and helps us to prioritize nodes with longer critical paths and nodes with many children. The actual algorithms used are somewhat more complex and are described in detail in dask/order.py
Initial Task Placement¶
When a new large batch of tasks come in and there are many idle workers then we want to give each worker a set of tasks that are close together/related and unrelated from the tasks given to other workers. This usually avoids inter-worker communication down the line. The same depth-first-with-child-weights priority given to workers described above can usually be used to properly segment the leaves of a graph into decently well separated sub-graphs with relatively low inter-sub-graph connectedness.
The last-in-first-out behavior used by the workers to minimize memory footprint can distort the task order provided by the clients. Tasks submitted recently may run sooner than tasks submitted long ago because they happen to be more convenient given the current data in memory. This behavior can be unfair but improves global runtimes and system efficiency, sometimes quite significantly.
However, workers inevitably run out of tasks that were related to tasks they were just working on and the last-in-first-out policy eventually exhausts itself. In these cases workers often pull tasks from the common task pool. The tasks in this pool are ordered in a first-come-first-served basis and so workers do behave in a fair scheduling manner at a coarse level if not a fine grained one.
Dask’s scheduling policies are short-term-efficient and long-term-fair.
Where these decisions are made¶
The objectives above are mostly followed by small decisions made by the client, scheduler, and workers at various points in the computation.
As we submit a graph from the client to the scheduler we assign a numeric priority to each task of that graph. This priority focuses on computing deeply before broadly, preferring critical paths, preferring nodes with many dependencies, etc.. This is the same logic used by the single-machine scheduler and lives in dask/order.py.
When the graph reaches the scheduler the scheduler changes each of these numeric priorities into a tuple of two numbers, the first of which is an increasing counter, the second of which is the client-generated priority described above. This per-graph counter encourages a first-in-first-out policy between computations. All tasks from a previous call to compute have a higher priority than all tasks from a subsequent call to compute (or submit, persist, map, or any operation that generates futures).
Whenever a task is ready to run the scheduler assigns it to a worker. The scheduler does not wait based on priority.
However when the worker receives these tasks it considers their priorities when determining which tasks to prioritize for communication or for computation. The worker maintains a heap of all ready-to-run tasks ordered by this priority.