Hao-Nan Zhu created HDFS-17627:
----------------------------------
Summary: Performance optimization on BlockUnderConstructionFeature
Key: HDFS-17627
URL: https://issues.apache.org/jira/browse/HDFS-17627
Project: Hadoop HDFS
Issue Type: Improvement
Components: server
Affects Versions: 3.3.0
Reporter: Hao-Nan Zhu
Hi, I’ve encountered performance bottlenecks in
_blockmanagement.BlockUnderConstructionFeature_ and I wonder if there's a
chance for optimization.
[_getStaleReplica()_|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockUnderConstructionFeature.java#L219]
may cause performance degradation when the list of replicas is large. The
method uses an *ArrayList* to collect stale replicas, which could cause memory
re-allocations and potential OOM errors when the number of stale replicas
increases. Furthermore,
[_getStaleReplica()_|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockUnderConstructionFeature.java#L219]
could also cause lock contention at some code paths like:
[_updatePipelineInternal()_|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/namenode/FSNamesystem.java#L6054]
(holding global lock) ->
[_updateLastBlock()_|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockManager.java#L4965]
->
[_setGenerationStampAndVerifyReplicas_|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockInfo.java#L426]{_}(){_}
->
[_getStaleReplica()_|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockUnderConstructionFeature.java#L219].
The optimization could be pre-sizing the ArrayList based on the actual number
of replicas (i.e. _List<ReplicaUnderConstruction> staleReplicas = new
ArrayList<>(replicas.length)_ ), which could minimize the number of times
resizing or reallocations. Another way to do the optimization is to have a
persisted list of {_}staleReplicas{_}, so there is no need to iterate over the
replicas.
Same issue could also happen with
[_appendUCPartsConcise()_|https://github.com/apache/hadoop/blob/6be04633b55bbd67c2875e39977cd9d2308dc1d1/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockUnderConstructionFeature.java#L349].
It takes in a StringBuilder with a [default size of
150|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/namenode/FSDirWriteFileOp.java#L792]
characters, which leads to risks of resizing when the number of replicas is
large. Within {_}BlockUnderConstructionFeature{_}, there are other similar
issues exist, including
[_addReplicaIfNotPresent()_|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockUnderConstructionFeature.java#L294]
or
[_setExpectedLocations()_|https://github.com/apache/hadoop/blob/2f0dd7c4feb1e482d47786d26d6d32483f39414b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockUnderConstructionFeature.java#L74].
Please let me know if there is something wrong with the analysis above, or any
comments on the optimization. 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]
