|
If you were logged in you would be able to see more operations.
|
|
|
|
File Attachments:
|
|
|
Environment:
|
All
|
|
Issue Links:
|
Reference
|
|
This issue relates to:
|
|
HADOOP-572
Chain reaction in a big cluster caused by simultaneous failure of only a few data-nodes.
|
|
|
|
|
|
|
|
| Resolution Date: |
20/Nov/06 11:25 PM
|
Currently the chooseTargets method (that selects datanodes for block-placement) takes in excess of 20% of cpu on a namenode. This is the most time-consuming namenode method, according to the profiler. This inefficiency has already contributed to cascading crash in DFS earlier. As datanodes went down, new locations needed to be found for the blocks on dead datanodes, and since this was done inside a synchronized method, it locked the whole namesystem for several minutes, which caused more datanode failures, when the namenode marked them dead because no heartbeat could be processed during that interval. This has been detailed in HADOOP-572.
The patch I am about to upload reduces the time taken in the chooseTarget method to be proportional to nReplicas per block, instead of the current implementation, which is proportional to (nDataNodes * nReplicas). Also, when a number of datanodes crash, their blocks are put on the pendingReplications list one datanode at a time in a synchronized section. (Currently, the syncchronized section processes ALL the dead datanodes, thus locking the namesystem for a considerable amount of time.) Also, this patch will add a unit test to check replication.
|
|
Description
|
Currently the chooseTargets method (that selects datanodes for block-placement) takes in excess of 20% of cpu on a namenode. This is the most time-consuming namenode method, according to the profiler. This inefficiency has already contributed to cascading crash in DFS earlier. As datanodes went down, new locations needed to be found for the blocks on dead datanodes, and since this was done inside a synchronized method, it locked the whole namesystem for several minutes, which caused more datanode failures, when the namenode marked them dead because no heartbeat could be processed during that interval. This has been detailed in HADOOP-572.
The patch I am about to upload reduces the time taken in the chooseTarget method to be proportional to nReplicas per block, instead of the current implementation, which is proportional to (nDataNodes * nReplicas). Also, when a number of datanodes crash, their blocks are put on the pendingReplications list one datanode at a time in a synchronized section. (Currently, the syncchronized section processes ALL the dead datanodes, thus locking the namesystem for a considerable amount of time.) Also, this patch will add a unit test to check replication. |
Show » |
|
Improvement
