hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Surendra Singh Lilhore (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HDFS-13768) Adding replicas to volume map makes DataNode start slowly
Date Tue, 18 Sep 2018 08:08:00 GMT

    [ https://issues.apache.org/jira/browse/HDFS-13768?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16618651#comment-16618651
] 

Surendra Singh Lilhore commented on HDFS-13768:
-----------------------------------------------

[~linyiqun] and [~arpitagarwal], I found some more improvment point.


 !screenshot-1.png!

If you see this screenshot {{FsDatasetUtil.getGenrationStampFromFile()}} and {{lock()}} is
taking so much time. This can be improved.

1. {{FsDatasetUtil.getGenrationStampFromFile()}}, This method loop over the all the file to
find out the meta file. This is time consuming.
 We can sort the list of file's and in sorted list next to block file metafile will be there,
so we no need to loop and find out the meta file.
{code:java}
  /**
   * Find the meta-file for the specified block file
   * and then return the generation stamp from the name of the meta-file.
   */
  static long getGenerationStampFromFile(List<File> files, File blockFile, int index)
      throws IOException {
    String blockName = blockFile.getName();
    if ((index + 1) < files.size()) {
      // Check if next index file is meta file
      String metaFile = files.get(index + 1).getName();
      if (metaFile.startsWith(blockName)) {
        return Block.getGenerationStamp(metaFile);
      }
    }
    //Search meta file in list 
    for (int j = 0; j < files.size(); j++) {
      ......
      ......
      return Block.getGenerationStamp(files.get(j).getName());
    }
    FsDatasetImpl.LOG.warn("Block " + blockFile + " does not have a metafile!");
    return HdfsConstants.GRANDFATHER_GENERATION_STAMP;
  }
{code}

After this change *60% datanode statup time* is redused in my cluster.

2. {{BlockPoolSlice.addReplicaToReplicasMap()}}, in this method first its find the oldReplica
from {{ReplicaMap}} and if it is null then it will add the ReplicaInfo in {{ReplicaMap}}.
For this its need to get the lock two time. Its better to add one method in ReplicaMap for
this work, so this work can be done in one lock. 

*Example:*
{code:java}
  /**
   * Add a replica's meta information into the map, if already exist return the
   * old replicaInfo
   * 
   * @param bpid
   * @param replicaInfo
   * @return
   */
  ReplicaInfo addAndGet(String bpid, ReplicaInfo replicaInfo) {
    checkBlockPool(bpid);
    checkBlock(replicaInfo);
    try (AutoCloseableLock l = lock.acquire()) {
      FoldedTreeSet<ReplicaInfo> set = map.get(bpid);
      if (set == null) {
        // Add an entry for block pool if it does not exist already
        set = new FoldedTreeSet<>();
        map.put(bpid, set);
      }
      ReplicaInfo oldReplicaInfo = set.get(replicaInfo.getBlockId(),
          LONG_AND_BLOCK_COMPARATOR);
      if (oldReplicaInfo != null) {
        return oldReplicaInfo;
      } else {
        set.add(replicaInfo);
      }
      return replicaInfo;
    }
  }
{code}

If you both are agree for this change then I will add this in next patch.

>  Adding replicas to volume map makes DataNode start slowly 
> -----------------------------------------------------------
>
>                 Key: HDFS-13768
>                 URL: https://issues.apache.org/jira/browse/HDFS-13768
>             Project: Hadoop HDFS
>          Issue Type: Bug
>    Affects Versions: 3.1.0
>            Reporter: Yiqun Lin
>            Assignee: Surendra Singh Lilhore
>            Priority: Major
>         Attachments: HDFS-13768.01.patch, HDFS-13768.02.patch, HDFS-13768.patch, screenshot-1.png
>
>
> We find DN starting so slowly when rolling upgrade our cluster. When we restart DNs,
the DNs start so slowly and not register to NN immediately. And this cause a lots of following
error:
> {noformat}
> DataXceiver error processing WRITE_BLOCK operation  src: /xx.xx.xx.xx:64360 dst: /xx.xx.xx.xx:50010
> java.io.IOException: Not ready to serve the block pool, BP-1508644862-xx.xx.xx.xx-1493781183457.
>         at org.apache.hadoop.hdfs.server.datanode.DataXceiver.checkAndWaitForBP(DataXceiver.java:1290)
>         at org.apache.hadoop.hdfs.server.datanode.DataXceiver.checkAccess(DataXceiver.java:1298)
>         at org.apache.hadoop.hdfs.server.datanode.DataXceiver.writeBlock(DataXceiver.java:630)
>         at org.apache.hadoop.hdfs.protocol.datatransfer.Receiver.opWriteBlock(Receiver.java:169)
>         at org.apache.hadoop.hdfs.protocol.datatransfer.Receiver.processOp(Receiver.java:106)
>         at org.apache.hadoop.hdfs.server.datanode.DataXceiver.run(DataXceiver.java:246)
>         at java.lang.Thread.run(Thread.java:745)
> {noformat}
> Looking into the logic of DN startup, it will do the initial block pool operation before
the registration. And during initializing block pool operation, we found the adding replicas
to volume map is the most expensive operation.  Related log:
> {noformat}
> 2018-07-26 10:46:23,771 INFO [Thread-105] org.apache.hadoop.hdfs.server.datanode.fsdataset.impl.FsDatasetImpl:
Time to add replicas to map for block pool BP-1508644862-xx.xx.xx.xx-1493781183457 on volume
/home/hard_disk/1/dfs/dn/current: 242722ms
> 2018-07-26 10:46:26,231 INFO [Thread-109] org.apache.hadoop.hdfs.server.datanode.fsdataset.impl.FsDatasetImpl:
Time to add replicas to map for block pool BP-1508644862-xx.xx.xx.xx-1493781183457 on volume
/home/hard_disk/5/dfs/dn/current: 245182ms
> 2018-07-26 10:46:32,146 INFO [Thread-112] org.apache.hadoop.hdfs.server.datanode.fsdataset.impl.FsDatasetImpl:
Time to add replicas to map for block pool BP-1508644862-xx.xx.xx.xx-1493781183457 on volume
/home/hard_disk/8/dfs/dn/current: 251097ms
> 2018-07-26 10:47:08,283 INFO [Thread-106] org.apache.hadoop.hdfs.server.datanode.fsdataset.impl.FsDatasetImpl:
Time to add replicas to map for block pool BP-1508644862-xx.xx.xx.xx-1493781183457 on volume
/home/hard_disk/2/dfs/dn/current: 287235ms
> {noformat}
> Currently DN uses independent thread to scan and add replica for each volume, but we
still need to wait the slowest thread to finish its work. So the main problem here is that
we could make the thread to run faster.
> The jstack we get when DN blocking in the adding replica:
> {noformat}
> "Thread-113" #419 daemon prio=5 os_prio=0 tid=0x00007f40879ff000 nid=0x145da runnable
[0x00007f4043a38000]
>    java.lang.Thread.State: RUNNABLE
> 	at java.io.UnixFileSystem.list(Native Method)
> 	at java.io.File.list(File.java:1122)
> 	at java.io.File.listFiles(File.java:1207)
> 	at org.apache.hadoop.fs.FileUtil.listFiles(FileUtil.java:1165)
> 	at org.apache.hadoop.hdfs.server.datanode.fsdataset.impl.BlockPoolSlice.addToReplicasMap(BlockPoolSlice.java:445)
> 	at org.apache.hadoop.hdfs.server.datanode.fsdataset.impl.BlockPoolSlice.addToReplicasMap(BlockPoolSlice.java:448)
> 	at org.apache.hadoop.hdfs.server.datanode.fsdataset.impl.BlockPoolSlice.addToReplicasMap(BlockPoolSlice.java:448)
> 	at org.apache.hadoop.hdfs.server.datanode.fsdataset.impl.BlockPoolSlice.getVolumeMap(BlockPoolSlice.java:342)
> 	at org.apache.hadoop.hdfs.server.datanode.fsdataset.impl.FsVolumeImpl.getVolumeMap(FsVolumeImpl.java:864)
> 	at org.apache.hadoop.hdfs.server.datanode.fsdataset.impl.FsVolumeList$1.run(FsVolumeList.java:191)
> {noformat}
> One improvement maybe we can use ForkJoinPool to do this recursive task, rather than
a sync way. This will be a great improvement because it can greatly speed up recovery process.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-help@hadoop.apache.org


Mime
View raw message