Home | About | Sematext search-lucene.com search-hadoop.com
 Search Hadoop and all its subprojects:

Switch to Plain View
Flume >> mail # user >> Lock contention in FileChannel


+
Pankaj Gupta 2013-08-13, 23:13
+
Hari Shreedharan 2013-08-13, 23:39
+
Pankaj Gupta 2013-08-14, 00:01
+
Hari Shreedharan 2013-08-14, 00:14
+
Brock Noland 2013-08-14, 00:51
+
Pankaj Gupta 2013-08-14, 02:06
Copy link to this message
-
Re: Lock contention in FileChannel
Even though the writes are done per batch, they don't go to disk rightaway - commits are the only ones which actually cause an fsync - which is when writes actually go to disk.  
Thanks,
Hari
On Tuesday, August 13, 2013 at 7:06 PM, Pankaj Gupta wrote:

> Looking at the code it seems like the lock and the i/o is done per event and not for a batch. Is that correct? If that is the case then it seems like there is a lot of overhead per event. The throughput I'm seeing is 1 - 1.5 MBps per disk which is way below the sequential read/write capacity of the disk which is easily over 50MBps. Adding more sinks doesn't help, they just block waiting for the queue to become free. CPU usage is 20%, there is enough RAM for page cache so that no read is going to disk. The queue seems to be the bottleneck. What is the throughput I should expect per disk?
>
>
> On Tue, Aug 13, 2013 at 5:51 PM, Brock Noland <[EMAIL PROTECTED] (mailto:[EMAIL PROTECTED])> wrote:
> > The lock is per file. Adding more directories to the channel will cause more files to be created. Of course you'll need additional disks behind those directories to see any performance increase.
> >
> >
> > On Tue, Aug 13, 2013 at 7:14 PM, Hari Shreedharan <[EMAIL PROTECTED] (mailto:[EMAIL PROTECTED])> wrote:
> > > Yes, IO is done inside locks to avoid multiple takes and puts getting written out at the same time. Even though Java makes sure the writes are serialized, Flume still needs to keep track of some counters etc, so the lock is required. Note that the lock you are talking about is  in the LogFile class, which represents a single file - so even if the write is inside that lock (which is also inside that class itself) that  does not cause any contention - because the lock is just preventing two IO ops to happen at the same time.
> > >
> > >
> > > Thanks,
> > > Hari
> > >
> > >
> > > On Tuesday, August 13, 2013 at 5:01 PM, Pankaj Gupta wrote:
> > >
> > > > It seems like some i/o is done inside the lock, which means that time for taking a lock is proportional to the time for i/o and thus it becomes a problem. I apologize in advance if I am wrong but the call stack and behavior I'm seeing seems to suggest that. Specifically, it seems that we do a write while inside take:
> > > > "SinkRunner-PollingRunner-LoadBalancingSinkProcessor" prio=10 tid=0x00007f857338c800 nid=0x404a runnable [0x00007f821b2f1000]
> > > >    java.lang.Thread.State: RUNNABLE
> > > >         at sun.nio.ch.NativeThread.current(Native Method)
> > > >         at sun.nio.ch.NativeThreadSet.add(NativeThreadSet.java:27)
> > > >         at sun.nio.ch.FileChannelImpl.write(FileChannelImpl.java:194)
> > > >         - locked <0x00000005190ec998> (a java.lang.Object)
> > > >         at org.apache.flume.channel.file.LogFile$Writer.write(LogFile.java:247)
> > > >         at org.apache.flume.channel.file.LogFile$Writer.take(LogFile.java:212)
> > > >         - locked <0x0000000519111590> (a org.apache.flume.channel.file.LogFileV3$Writer)
> > > >         at org.apache.flume.channel.file.Log.take(Log.java:550)
> > > >         at org.apache.flume.channel.file.FileChannel$FileBackedTransaction.doTake(FileChannel.java:499)
> > > >         at org.apache.flume.channel.BasicTransactionSemantics.take(BasicTransactionSemantics.java:113)
> > > >         at org.apache.flume.channel.BasicChannelSemantics.take(BasicChannelSemantics.java:95)
> > > >         at org.apache.flume.sink.AbstractRpcSink.process(AbstractRpcSink.java:330)
> > > >         at org.apache.flume.sink.LoadBalancingSinkProcessor.process(LoadBalancingSinkProcessor.java:154)
> > > >         at org.apache.flume.SinkRunner$PollingRunner.run(SinkRunner.java:147)
> > > >         at java.lang.Thread.run(Thread.java:662)
> > > >
> > > >
> > > >
> > > > On Tue, Aug 13, 2013 at 4:39 PM, Hari Shreedharan <[EMAIL PROTECTED] (mailto:[EMAIL PROTECTED])> wrote:
> > > > > Since the channel is designed to make sure that events are not duplicated to multiple sinks, and to protect against corruption due to concurrency issues, we do not need the locking in the channel's flume event queue. It is unlikely that locking is what is causing performance issues because the channel is heavily I/O bound. If you take a series of thread dumps, you will probably see that those threads are moving forward and the ones reading/writing from/to disk are the ones which are slower. These locks are unlikely to hit performance much.
+
Brock Noland 2013-08-14, 02:22
+
Pankaj Gupta 2013-08-14, 02:33
+
Brock Noland 2013-08-14, 02:41
+
Pankaj Gupta 2013-08-14, 02:46
+
Brock Noland 2013-08-14, 02:54
+
Pankaj Gupta 2013-08-14, 02:57
+
Brock Noland 2013-08-14, 03:06
+
Pankaj Gupta 2013-08-14, 03:16
+
Brock Noland 2013-08-14, 03:30
+
Pankaj Gupta 2013-08-14, 18:57
+
Pankaj Gupta 2013-08-14, 19:12
+
Pankaj Gupta 2013-08-14, 19:34
+
Hari Shreedharan 2013-08-14, 19:43
+
Pankaj Gupta 2013-08-14, 19:59
+
Pankaj Gupta 2013-08-15, 06:04
+
Pankaj Gupta 2013-08-18, 04:43
+
Hari Shreedharan 2013-08-14, 19:04
+
Pankaj Gupta 2013-08-14, 02:16