[HDFS-4130] BKJM: The reading for editlog at NN starting using bkjm is not efficient

• Type: Sub-task
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.0.2-alpha, 3.0.0-alpha1
• Fix Version/s: 2.0.3-alpha
• Component/s: ha, performance
• Target Version/s: 2.0.3-alpha
• Hadoop Flags: Reviewed
• Assignee: Han Xiao
• Reporter: Han Xiao
• Created: 30/Oct/12 09:01
• Updated: 12/May/16 18:12
• Resolved: 10/Dec/12 18:57

Description

Now, the method of BookKeeperJournalManager.selectInputStreams is written like:

while (true) {
EditLogInputStream elis;
try

{ elis = getInputStream(fromTxId, inProgressOk); }

catch (IOException e)

{ LOG.error(e); return; }

if (elis == null)

{ return; }
streams.add(elis);
if (elis.getLastTxId() == HdfsConstants.INVALID_TXID) { return; }

fromTxId = elis.getLastTxId() + 1;
}

EditLogInputstream is got from getInputStream(), which will read the ledgers from zookeeper in each calling.
This will be a larger cost of times when the the number ledgers becomes large.
The reading of ledgers from zk is not necessary for every calling of getInputStream().

The log of time wasting here is as follows:
2012-10-30 16:44:52,995 INFO org.apache.hadoop.hdfs.server.namenode.NameNode: Caching file names occuring more than 10 times
2012-10-30 16:49:24,643 INFO hidden.bkjournal.org.apache.bookkeeper.proto.PerChannelBookieClient: Successfully connected to bookie: /167.52.1.121:318

The stack of the process when blocking between the two lines of log is like:
"main" prio=10 tid=0x000000004011f000 nid=0x39ba in Object.wait() [0x00007fca020fe000]
java.lang.Thread.State: WAITING (on object monitor)
at java.lang.Object.wait(Native Method)
at java.lang.Object.wait(Object.java:485)
at hidden.bkjournal.org.apache.zookeeper.ClientCnxn.submitRequest(ClientCnxn.java:1253)
- locked <0x00000006fb8495a8> (a hidden.bkjournal.org.apache.zookeeper.ClientCnxn$Packet)
at hidden.bkjournal.org.apache.zookeeper.ZooKeeper.getData(ZooKeeper.java:1129)
at org.apache.hadoop.contrib.bkjournal.utils.RetryableZookeeper.getData(RetryableZookeeper.java:501)
at hidden.bkjournal.org.apache.zookeeper.ZooKeeper.getData(ZooKeeper.java:1160)
at org.apache.hadoop.contrib.bkjournal.EditLogLedgerMetadata.read(EditLogLedgerMetadata.java:113)
at org.apache.hadoop.contrib.bkjournal.BookKeeperJournalManager.getLedgerList(BookKeeperJournalManager.java:725)
at org.apache.hadoop.contrib.bkjournal.BookKeeperJournalManager.getInputStream(BookKeeperJournalManager.java:442)
at org.apache.hadoop.contrib.bkjournal.BookKeeperJournalManager.selectInputStreams(BookKeeperJournalManager.java:480)

betweent different time, the diff of stack is:
diff stack stack2
1c1
< 2012-10-30 16:44:53
—
> 2012-10-30 16:46:17
106c106
< - locked <0x00000006fb8495a8> (a hidden.bkjournal.org.apache.zookeeper.ClientCnxn$Packet)
—
> - locked <0x00000006fae58468> (a hidden.bkjournal.org.apache.zookeeper.ClientCnxn$Packet)

In our environment, the waiting time could even reach to tens of minutes.

Issue Links

• breaks: HDFS-4445 All BKJM ledgers are not checked while tailing, So failover will fail.