[
https://issues.apache.org/jira/browse/HDFS-13671?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16511402#comment-16511402
]
Daryn Sharp commented on HDFS-13671:
------------------------------------
I'm not surprised. I fully expected having to revert when we start scale
testing 3.x.
The folded tree jira has micro-benchmarks demonstrating a 4x degradation in
performance. Micro-benchmarks tend to exaggerate performance differences, so
let's see how that measures up in the real world.
Baseline is a large delete from last month on a 2.8 cluster:
* 29.9M blocks; 87 seconds = 344k blocks/sec
Let's analyze permutations of the reported numbers:
* 14M blocks; 3 min = 78k blocks/sec = 4X slower
* 12M blocks; 3 min = 67k blocks/sec = 5X slower
* 14M blocks; 6 min = 39k blocks/sec = 9X slower
It appears it's _at least as bad_ as predicted by the micro-benchmarks. The
tree traversal, the internal node copies, and the rebalancing aren't cheap.
> Namenode deletes large dir slowly caused by FoldedTreeSet#removeAndGet
> ----------------------------------------------------------------------
>
> Key: HDFS-13671
> URL: https://issues.apache.org/jira/browse/HDFS-13671
> Project: Hadoop HDFS
> Issue Type: Bug
> Affects Versions: 3.1.0, 3.0.3
> Reporter: Yiqun Lin
> Priority: Major
>
> NameNode hung when deleting large files/blocks. The stack info:
> {code}
> "IPC Server handler 4 on 8020" #87 daemon prio=5 os_prio=0
> tid=0x00007fb505b27800 nid=0x94c3 runnable [0x00007fa861361000]
> java.lang.Thread.State: RUNNABLE
> at
> org.apache.hadoop.hdfs.util.FoldedTreeSet.compare(FoldedTreeSet.java:474)
> at
> org.apache.hadoop.hdfs.util.FoldedTreeSet.removeAndGet(FoldedTreeSet.java:849)
> at
> org.apache.hadoop.hdfs.util.FoldedTreeSet.remove(FoldedTreeSet.java:911)
> at
> org.apache.hadoop.hdfs.server.blockmanagement.DatanodeStorageInfo.removeBlock(DatanodeStorageInfo.java:252)
> at
> org.apache.hadoop.hdfs.server.blockmanagement.BlocksMap.removeBlock(BlocksMap.java:194)
> at
> org.apache.hadoop.hdfs.server.blockmanagement.BlocksMap.removeBlock(BlocksMap.java:108)
> at
> org.apache.hadoop.hdfs.server.blockmanagement.BlockManager.removeBlockFromMap(BlockManager.java:3813)
> at
> org.apache.hadoop.hdfs.server.blockmanagement.BlockManager.removeBlock(BlockManager.java:3617)
> at
> org.apache.hadoop.hdfs.server.namenode.FSNamesystem.removeBlocks(FSNamesystem.java:4270)
> at
> org.apache.hadoop.hdfs.server.namenode.FSNamesystem.deleteInternal(FSNamesystem.java:4244)
> at
> org.apache.hadoop.hdfs.server.namenode.FSNamesystem.deleteInt(FSNamesystem.java:4180)
> at
> org.apache.hadoop.hdfs.server.namenode.FSNamesystem.delete(FSNamesystem.java:4164)
> at
> org.apache.hadoop.hdfs.server.namenode.NameNodeRpcServer.delete(NameNodeRpcServer.java:871)
> at
> org.apache.hadoop.hdfs.server.namenode.AuthorizationProviderProxyClientProtocol.delete(AuthorizationProviderProxyClientProtocol.java:311)
> at
> org.apache.hadoop.hdfs.protocolPB.ClientNamenodeProtocolServerSideTranslatorPB.delete(ClientNamenodeProtocolServerSideTranslatorPB.java:625)
> at
> org.apache.hadoop.hdfs.protocol.proto.ClientNamenodeProtocolProtos$ClientNamenodeProtocol$2.callBlockingMethod(ClientNamenodeProtocolProtos.java)
> at
> org.apache.hadoop.ipc.ProtobufRpcEngine$Server$ProtoBufRpcInvoker.call(ProtobufRpcEngine.java:617)
> {code}
> In the current deletion logic in NameNode, there are mainly two steps:
> * Collect INodes and all blocks to be deleted, then delete INodes.
> * Remove blocks chunk by chunk in a loop.
> Actually the first step should be a more expensive operation and will takes
> more time. However, now we always see NN hangs during the remove block
> operation.
> Looking into this, we introduced a new structure {{FoldedTreeSet}} to have a
> better performance in dealing FBR/IBRs. But compared with early
> implementation in remove-block logic, {{FoldedTreeSet}} seems more slower
> since It will take additional time to balance tree node. When there are large
> block to be removed/deleted, it looks bad.
> For the get type operations in {{DatanodeStorageInfo}}, we only provide the
> {{getBlockIterator}} to return blocks iterator and no other get operation
> with specified block. Still we need to use {{FoldedTreeSet}} in
> {{DatanodeStorageInfo}}? As we know {{FoldedTreeSet}} is benefit for Get not
> Update. Maybe we can revert this to the early implementation.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
