Current synchronization on FSDataset seems not quite right. Doing what amounted
to applying Todd's patch that modifies FSDataSet to use reentrant rwlocks
cleared up that type of problem for us.
- Andy
From: Jeff Whiting <je...@qualtrics.com>
Subject: Re: Lots of Different Kind of Datanode Errors
To: hdfs-user@hadoop.apache.org
Date: Monday, June 7, 2010, 10:02 AM
Thanks for the replies. I have turned off swap on all the machines to
prevent any swap problems. I was pounding my hard drives quite hard.
I had a simulated 60 clients loading data as fast as I could into hbase
with a map reduce export job going at the same time. Would that
scenario explain some of the errors I was seeing?
Over the weekend under more of a normal load I haven't not any
exception except for about 6 of these:
2010-06-05 03:46:41,229 ERROR datanode.DataNode
(DataXceiver.java:run(131)) - DatanodeRegistration(192.168.0.98:50010,
storageID=DS-1806250311-192.168.0.98-50010-1274208294562,
infoPort=50075, ipcPort=50020):DataXceiver
org.apache.hadoop.hdfs.server.datanode.BlockAlreadyExistsException:
Block blk_-1677111232590888964_4471547 is valid, and cannot be written
to.
at
org.apache.hadoop.hdfs.server.datanode.FSDataset.writeToBlock(FSDataset.java:999)
The reason the config shows 4096 is because I increased the xceiver
account after the first email message in this thread.
~Jeff
Allen Wittenauer wrote:
On Jun 4, 2010, at 12:03 PM, Todd Lipcon wrote:
Hi Jeff,
That seems like a reasonable config, but the error message you pasted indicated
xceivers was set to 2048 instead of 4096.
Also, in my experience SocketTimeoutExceptions are usually due to swapping.
Verify that your machines aren't swapping when you're under load.
Or doing any other heavy disk IO.
--
Jeff Whiting
Qualtrics Senior Software Engineer
je...@qualtrics.com
