[
https://issues.apache.org/jira/browse/HADOOP-2960?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12576057#action_12576057
]
Owen O'Malley commented on HADOOP-2960:
---------------------------------------
Computing the number of unique keys is not free. In particular, it will cost
O(N) comparisons. Even worse, this doesn't scale. Currently, combiners are only
applied on the original spill, where your approach could be done. However, we
plan to apply combiners every time you write to disk during the merge sort.
There, you certainly can't count the duplicated keys without a prohibitive
cost.
-1
Once we have HADOOP-2399, almost any reduction in the cost of network and disk
i/o should be worth the cost of the combiner.
> A mapper should use some heuristics to decide whether to run the combiner
> during spills
> ---------------------------------------------------------------------------------------
>
> Key: HADOOP-2960
> URL: https://issues.apache.org/jira/browse/HADOOP-2960
> Project: Hadoop Core
> Issue Type: Bug
> Reporter: Runping Qi
>
> Right now, the combiner, if set, will be called for each spill, no mapper
> whether the combiner can actually reduce the values.
> The mapper should use some heuristics to decide whether to run the combiner
> during spills.
> One of such heuristics is to check the the ratio of the nymber of keys to
> the number of unique keys in the spill.
> The combiner will be called only if that ration exceeds certain threshold
> (say 2).
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
