Details
-
Improvement
-
Status: Closed
-
Major
-
Resolution: Duplicate
-
0.2.0
-
None
-
None
-
large clusters
Description
When dfs.data.dir has multiple values, we currently start a DataNode for each (all in the same JVM). Instead we should run a single DataNode that stores block files into the different directories. This will reduce the number of connections to the namenode. We cannot hash because different devices might be different amounts full. So the datanode will need to keep a table mapping from block id to file location, and add new blocks to less full devices.
Attachments
Issue Links
- is duplicated by
-
HADOOP-64 DataNode should be capable of managing multiple volumes
- Closed