Work Stealing

Some tasks prefer to run on certain workers. This may be because that worker holds data dependencies of the task or because the user has expressed a loose desire that the task run in a particular place. Occasionally this results in a few very busy workers and several idle workers. In this situation the idle workers may choose to steal work from the busy workers, even if stealing work requires the costly movement of data.

This is a performance optimization and not required for correctness. Work stealing provides robustness in many ad-hoc cases, but can also backfire when we steal the wrong tasks and reduce performance.

Criteria for stealing

Computation to Communication Ratio

Stealing is profitable when the computation time for a task is much longer than the communication time of the task’s dependencies.

Bad example

We do not want to steal tasks that require moving a large dependent piece of data across a wire from the victim to the thief if the computation is fast. We end up spending far more time in communication than just waiting a bit longer and giving up on parallelism.

[data] = client.scatter([np.arange(1000000000)])
x = client.submit(np.sum, data)

Good example

We do want to steal task tasks that only need to move dependent pieces of data, especially when the computation time is expensive (here 100 seconds.)

[data] = client.scatter([100])
x = client.submit(sleep, data)

Fortunately we often know both the number of bytes of dependencies (as reported by calling sys.getsizeof on the workers) and the runtime cost of previously seen functions, which is maintained as an exponentially weighted moving average.

Saturated Worker Burden

Stealing may be profitable even when the computation-time to communication-time ratio is poor. This occurs when the saturated workers have a very long backlog of tasks and there are a large number of idle workers. We determine if it acceptable to steal a task if the last task to be run by the saturated workers would finish more quickly if stolen or if it remains on the original/victim worker.

The longer the backlog of stealable tasks, and the smaller the number of active workers we have both increase our willingness to steal. This is balanced against the compute-to-communicate cost ratio.

Steal from the Rich

We would like to steal tasks from particularly over-burdened workers rather than workers with just a few excess tasks.


If a task has been specifically restricted to run on particular workers (such as is the case when special hardware is required) then we do not steal.

Choosing tasks to steal

We maintain a list of sets of stealable tasks, ordered into bins by computation-to-communication time ratio. The first bin contains all tasks with a compute-to-communicate ratio greater than or equal to 8 (considered high enough to always steal), the next bin with a ratio of 4, the next bin with a ratio of 2, etc.., all the way down to a ratio of 1/256, which we will never steal.

This data structure provides a somewhat-ordered view of all stealable tasks that we can add to and remove from in constant time, rather than log(n) as with more traditional data structures, like a heap.

During any stage when we submit tasks to workers we check if there are both idle and saturated workers and if so we quickly run through this list of sets, selecting tasks from the best buckets first, working our way down to the buckets of less desirable stealable tasks. We stop either when there are no more stealable tasks, no more idle workers, or when the quality of the task-to-be-stolen is not high enough given the current backlog.

This approach is fast, optimizes to steal the tasks with the best computation-to-communication cost ratio (up to a factor of two) and tends to steal from the workers that have the largest backlogs, just by nature that random selection tends to draw from the largest population.