Details
-
Improvement
-
Status: Open
-
Major
-
Resolution: Unresolved
-
None
-
None
-
None
-
None
Description
As far as possible, configurations should be independent of cluster size. The cluster size is a parameter that the system knows and should thus be used to size things accordingly. Currently we know that the following parameters must be adjusted to non-default values for large clusters:
- mapred.job.tracker.handler.count
- mapred.reduce.parallel.copies
- tasktracker.http.threads
We should attempt to make each of these either set to something proportional to cluster size or (harder) dynamically sized based on load.