mirror of
https://github.com/vale981/ray
synced 2025-03-08 19:41:38 -05:00
26 lines
1.3 KiB
Markdown
26 lines
1.3 KiB
Markdown
|
# Scheduler
|
||
|
|
|
||
|
|
The scheduling strategies currently implemented in Photon are fairly basic and
|
||
|
|
all use a central scheduler.
|
||
|
|
|
||
|
|
* The naive scheduler assigns tasks to workers just taking into account
|
||
|
|
dependencies between tasks (no other information like data locality). It is
|
||
|
|
supposed to be an example for how to write a scheduler. We do not recommend
|
||
|
|
its use and it only works well for single node setups. Tasks are assigned in the
|
||
|
|
following way: For each idle worker, we iterate over the tasks in the task
|
||
|
|
queue. The first task that has all its requirements satisfied will be scheduled
|
||
|
|
on the worker.
|
||
|
|
|
||
|
|
* The locality aware scheduler is more suited for multi node setups, but still
|
||
|
|
inappropriate for very large clusters. This is because the computational
|
||
|
|
overhead for each scheduling decision is O(mn) where m is the number of idle
|
||
|
|
workers and n is the number of tasks in the task queue. For each idle worker,
|
||
|
|
all tasks in the task queue are considered and the one that requires the
|
||
|
|
smallest number of objects to be shipped will be executed.
|
||
|
|
|
||
|
|
We expect to implement more refined scheduling strategies in the future,
|
||
|
|
including more computationally efficient location aware scheduling,
|
||
|
|
scheduling that takes into account sizes of the shipped objects, and strategies
|
||
|
|
that do not require a central scheduler (which is a bottleneck for large
|
||
|
|
clusters).
|