Github user jkbradley commented on the pull request:
https://github.com/apache/spark/pull/2868#issuecomment-61167709
I ran some local tests but did not see any speedups. This was trying to
mimic your earlier test:
* original mnist dataset
* depths 5, 10, 20, and 30
* 1 compute node (figuring 1 node would make node ID caching most helpful)
Running times were virtually the same between the master & the master +
this PR. I only tested it with 1 tree and 5 trees, but I would not expect that
to make a huge difference from 100 trees.
Thinking about when distributed node ID caching might speed things up, I
could imagine it being most helpful with an imbalanced tree, where there was
some path which a lot of instances followed. I.e., the tree gets very deep
before we can switch to local training. This sounds a bit unlikely to me.
However, I would be OK with merging this as protection against such an event,
especially since I did not see slowdowns from this PR.
What do you think?
---
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.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
