Chenyu Zheng created HDFS-17542:
-----------------------------------
Summary: EC: Optimize the EC block reconstruction.
Key: HDFS-17542
URL: https://issues.apache.org/jira/browse/HDFS-17542
Project: Hadoop HDFS
Issue Type: Improvement
Reporter: Chenyu Zheng
Assignee: Chenyu Zheng
The current reconstruction process of EC blocks is based on the original
contiguous blocks. It is mainly implemented through the work constructed by
computeReconstructionWorkForBlocks. It can be roughly divided into three
processes:
* scheduleReconstruction
* chooseTargets
* validateReconstructionWork
For ordinary contiguous blocks:
* (1) scheduleReconstruction
Select srcNodes as the source of the copy block according to the status of each
replica of the block.
* (2) chooseTargets
Select the target of the copy.
* (3) validateReconstructionWork
Add the copy command to srcNode, srcNode receives the command through
heartbeat, and executes the block copy from src to target.
For EC blocks:
(1) and (2) are nearly same. However, in (3), block copying or block
reconstruction may occur, or no work may be generated, such as when some
storage are busy. If no work is generated, it will lead to the problem
described in HDFS-17516. Even if no block copying or block reconstruction is
generated, pendingReconstruction and neededReconstruction will still be updated
until the block times out, which wastes the scheduling opportunity.
In order to be compatible with the original contiguous blocks and decide the
specific action in (3), unnecessary liveBlockIndices, liveBusyBlockIndices, and
excludeReconstructedIndices are introduced. We know many bug is related here.
These can be avoided.
Improvements:
* Move the work of deciding whether to copy or reconstruct blocks from (3) to
(1).
Such improvements are more conducive to implementing the explicit specification
of the reconstruction block index mentioned in HDFS-16874, and do not need to
pass liveBlockIndices, liveBusyBlockIndice.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
