Goodness Ayinmode created HDFS-17638:
----------------------------------------
Summary: Lock contention for DatanodeStorageInfo when the number
of storage nodes is large
Key: HDFS-17638
URL: https://issues.apache.org/jira/browse/HDFS-17638
Project: Hadoop HDFS
Issue Type: Improvement
Components: datanode, server
Affects Versions: 3.4.0
Reporter: Goodness Ayinmode
Hi,
I was looking into the DatanodeStorageInfo class and I think some of the
methods could give issues at large scale. For example, to convert
DatanodeStorageInfo objects into their respective DatanodeDescriptor and
Storage ID forms,
[DatanodeStorageInfo.toDatanodeInfos()|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/DatanodeStorageInfo.java#L44]
and
[DatanodeStorageInfo.toStorageIDs()|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/DatanodeStorageInfo.java#L61]
iterate over the entire array of storage nodes. Each operation is linear,
however performance issues can arise, when they are called under a lock, like
in
[bumpBlockGenerationStamp|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/namenode/FSNamesystem.java#L5987],
where
[newLocatedBlock|https://github.com/apache/hadoop/blob/49a495803a9451850b8982317e277b605c785587/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockManager.java#L5461]
calls both methods (bumpBlockGenerationStamp --> newLocatedBlock -->
[newLocatedBlock|https://github.com/apache/hadoop/blob/49a495803a9451850b8982317e277b605c785587/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockManager.java#L5437]
or
[newLocatedStripedBlock|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockManager.java#L5435]
--> toDatanodeInfos and toStorageIDs) under the writeLock. This situation can
be even more problematic when these methods are repeatedly invoked within an
iteration like in
[createLocatedBlockList|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockManager.java#L1450]
([createLocatedBlocks|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockManager.java#L1601]
--> createLocatedBlockList --> [createLocatedBlock
|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockManager.java#L1487]-->
newLocatedBlock or newLocatedStripedBlock --> toDatanodeInfos and
toStorageIDs). Such behaviors cause significant synchronization bottlenecks
when the number of blocks or number of storage nodes is large.
[BlockPlacementPolicyDefault.getPipeline|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockPlacementPolicyDefault.java#L1147],
[BlockPlacementPolicyDefault.chooseTarget|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockPlacementPolicyDefault.java#L287],
and
[BlockManager.validateReconstructionWork|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockManager.java#L2355]
(
[BlockManager.computeReconstructionWorkForBlocks|https://github.com/apache/hadoop/blob/49a495803a9451850b8982317e277b605c785587/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockManager.java#L2187]
--> BlockManager.validateReconstructionWork -->
[incrementBlocksScheduled|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/DatanodeStorageInfo.java#L338]
) also faces a similar issue with lock contention.
Please let me know if my analysis is wrong, and if there are suggestions to
make this better. Thanks
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
