Github user mridulm commented on the pull request:
https://github.com/apache/spark/pull/1313#issuecomment-49207097
My rationale for suggesting noPrefs tasks after NODE_LOCAL was to ensure
that noPref tasks do not preempt locality for NODE_LOCAL tasks (they cant for
PROCESS_LOCAL)
I think noPrefs sort of breaks the current model where we wait for
PROCESS_LOCAL, then NODE_LOCAL and so on : though I dont seem to have a good
solution.
For example: if we have 20 executors, with 30 no pref tasks and 100
NODE_LOCAL tasks.
Depending on the time taken for the no pref tasks, we can end up with
most/none of the NODE_LOCAL tasks actually getting scheduled as NODE_LOCAL
(since the corresponding executors might be running no pref tasks). This is
slightly contrived :-) But it does have implications in general scheduling too.
I did not properly account for idle executors - which leads to suboptimal
utilization in my suggestion above.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
