Hi Jason,
2 million blocks per data-node is not going to work.
There were discussions about it previously, please
check the mail archives.
This means you have a lot of very small files, which
HDFS is not designed to support. A general recommendation
is to group small files into large ones, introducing
some kind of record structure delimiting those small files,
and control it in on the application level.
Thanks,
--Konstantin
Jason Venner wrote:
The problem we are having is that datanodes periodically stall for 10-15
minutes and drop off the active list and then come back.
What is going on is that a long operation set is holding the lock on on
FSDataset.volumes, and all of the other block service requests stall
behind this lock.
"DataNode: [/data/dfs-video-18/dfs/data]" daemon prio=10 tid=0x4d7ad400
nid=0x7c40 runnable [0x4c698000..0x4c6990d0]
java.lang.Thread.State: RUNNABLE
at java.lang.String.lastIndexOf(String.java:1628)
at java.io.File.getName(File.java:399)
at
org.apache.hadoop.dfs.FSDataset$FSDir.getGenerationStampFromFile(FSDataset.java:148)
at
org.apache.hadoop.dfs.FSDataset$FSDir.getBlockInfo(FSDataset.java:181)
at
org.apache.hadoop.dfs.FSDataset$FSVolume.getBlockInfo(FSDataset.java:412)
at
org.apache.hadoop.dfs.FSDataset$FSVolumeSet.getBlockInfo(FSDataset.java:511)
- locked <0x551e8d48> (a org.apache.hadoop.dfs.FSDataset$FSVolumeSet)
at org.apache.hadoop.dfs.FSDataset.getBlockReport(FSDataset.java:1053)
at org.apache.hadoop.dfs.DataNode.offerService(DataNode.java:708)
at org.apache.hadoop.dfs.DataNode.run(DataNode.java:2890)
at java.lang.Thread.run(Thread.java:619)
This is basically taking a stat on every hdfs block on the datanode,
which in our case is ~ 2million, and can take 10+ minutes (we may be
experiencing problems with our raid controller but have no visibility
into it) at the OS level the file system seems fine and operations
eventually finish.
It appears that a couple of different data structures are being locked
with the single object FSDataset$Volume.
Then this happens:
"org.apache.hadoop.dfs.datanode$dataxcei...@1bcee17" daemon prio=10
tid=0x4da8d000 nid=0x7ae4 waiting for monitor entry
[0x459fe000..0x459ff0d0]
java.lang.Thread.State: BLOCKED (on object monitor)
at
org.apache.hadoop.dfs.FSDataset$FSVolumeSet.getNextVolume(FSDataset.java:473)
- waiting to lock <0x551e8d48> (a
org.apache.hadoop.dfs.FSDataset$FSVolumeSet)
at org.apache.hadoop.dfs.FSDataset.writeToBlock(FSDataset.java:934)
- locked <0x54e550e0> (a org.apache.hadoop.dfs.FSDataset)
at
org.apache.hadoop.dfs.DataNode$BlockReceiver.<init>(DataNode.java:2322)
at
org.apache.hadoop.dfs.DataNode$DataXceiver.writeBlock(DataNode.java:1187)
at org.apache.hadoop.dfs.DataNode$DataXceiver.run(DataNode.java:1045)
at java.lang.Thread.run(Thread.java:619)
which locks the FSDataset while waiting on the volume object
and now all of the Datanode operations stall waiting on the FSDataset
object.
----------
Our particular installation doesn't use multiple directories for hdfs,
so a first simple hack for a local fix would be to modify getNextVolume
to just return the single volume and not be synchronized
A richer alternative would be to make the locking more fine grained on
FSDataset$FSVolumeSet.
Of course we are also trying to fix the file system performance and dfs
block loading that results in the block report taking a long time.
Any suggestions or warnings?
Thanks.
