Closed lgray closed 7 months ago
This results in a task graph with better parallelism since now the leaves of the tree reduce do not wait on each other outside of split_every sized blocks. Better fault recovery too (fewer recomputed subgraphs if there's a failure).
split_every
This results in a task graph with better parallelism
For posterity, can we have images of before and after?
Here's before:
Here's after:
This results in a task graph with better parallelism since now the leaves of the tree reduce do not wait on each other outside of
split_every
sized blocks. Better fault recovery too (fewer recomputed subgraphs if there's a failure).