[
https://issues.apache.org/jira/browse/CASSANDRA-11818?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15295611#comment-15295611
]
Robert Stupp edited comment on CASSANDRA-11818 at 5/22/16 3:59 PM:
-------------------------------------------------------------------
This thing's getting weird. I tried a couple of different combinations:
# with unpooled direct buffers (change in {{CBUtil}}: {{allocator = new
UnpooledByteBufAllocator(true)}}) against Java 1.8.0_92
# with unpooled heap buffers (change in {{CBUtil}}: {{allocator = new
UnpooledByteBufAllocator(false)}}) against Java 1.8.0_92
Variant 1 showed the same behavior. Looks like GC causes long STW, causing
responses to pile up and then trying to allocate a totally huge amount of
direct memory
Variant 2 showed a strange behavior with PS GC _and_ ParNew/CMS GC: Works fine
until the old gen becomes eligible for GC. And then ends with an effectively
endless-GC-loop. (CMS basically blocked the application according to GC log)
That led me to the idea to try it with Java 1.8.0_66 - and that works fine with
variant 2. Java 1.8.0_77 also works fine with variant 2. So it looks like Java
1.8.0_92 is somehow broken. (Didn’t try variant 1)
Unfortunately, the default variant with pooled byte buffers does not work with
1.8.0_77. It shows the same behavior as with u92.
So, the underlying problem remains: if GC runs wild (i.e. runs into some longer
STW phases - some hundred milliseconds in my test), requests pile up and cause
a lot of allocations, which in turn can cause OOMs. I have no idea why it runs
into these many, consecutive, long GCs. Just see the result: lots of allocated
direct memory, that’s never freed. It then looks like as in the following
screenshot. At approx 16:03 a couple of ParNews with 200-500ms kicked in but
stopped after a while. At 16:06 some ParNews with 200-500ms STW occurred and
lasted until the load driver was killed. As a result all (allowed) direct
memory is allocated (configured with 2g max direct memory).
pooled direct mem usage w/ 1.8.0u77:
!11818-direct-mem.png|width=1000!
Further, using unpooled direct buffers work a bit better than pooled direct
buffers: At least, the unpooled buffers are released after a couple of minutes
and C* is responsive again.
unpooled direct mem usage w/ 1.8.0u77:
!11818-direct-mem-unpooled.png|width=1000!
The buffer allocations that cause this problem are for messages with around 4MB
or more. Could be, that it is probably not a good idea to use a pre-thread
arena for messages that are that big.
/cc [~norman], do you have any idea?
was (Author: snazy):
This thing's getting weird. I tried a couple of different combinations:
# with unpooled direct buffers (change in {{CBUtil}}: {{allocator = new
UnpooledByteBufAllocator(true)}}) against Java 1.8.0_92
# with unpooled heap buffers (change in {{CBUtil}}: {{allocator = new
UnpooledByteBufAllocator(false)}}) against Java 1.8.0_92
Variant 1 showed the same behavior. Looks like GC causes long STW, causing
responses to pile up and then trying to allocate a totally huge amount of
direct memory
Variant 2 showed a strange behavior with PS GC _and_ ParNew/CMS GC: Works fine
until the old gen becomes eligible for GC. And then ends with an effectively
endless-GC-loop. (CMS basically blocked the application according to GC log)
That led me to the idea to try it with Java 1.8.0_66 - and that works fine with
variant 2. Java 1.8.0_77 also works fine with variant 2. So it looks like Java
1.8.0_92 is somehow broken. (Didn’t try variant 1)
Unfortunately, the default variant with pooled byte buffers does not work with
1.8.0_77. It shows the same behavior as with u92.
So, the underlying problem remains: if GC runs wild (i.e. runs into some longer
STW phases - some hundred milliseconds in my test), requests pile up and cause
a lot of allocations, which in turn can cause OOMs. I have no idea why it runs
into these many, consecutive, long GCs. Just see the result: lots of allocated
direct memory, that’s never freed. It then looks like as in the following
screenshot. At approx 16:03 a couple of ParNews with 200-500ms kicked in but
stopped after a while. At 16:06 some ParNews with 200-500ms STW occurred and
lasted until the load driver was killed. As a result all (allowed) direct
memory is allocated (configured with 2g max direct memory).
!11818-direct-mem.png|pooled direct mem usage w/ 1.8.0u77!
Further, using unpooled direct buffers work a bit better than pooled direct
buffers: At least, the unpooled buffers are released after a couple of minutes
and C* is responsive again.
!11818-direct-mem-unpooled.png|unpooled direct mem usage w/ 1.8.0u77!
The buffer allocations that cause this problem are for messages with around 4MB
or more. Could be, that it is probably not a good idea to use a pre-thread
arena for messages that are that big.
/cc [~norman], do you have any idea?
> C* does neither recover nor trigger stability inspector on direct memory OOM
> ----------------------------------------------------------------------------
>
> Key: CASSANDRA-11818
> URL: https://issues.apache.org/jira/browse/CASSANDRA-11818
> Project: Cassandra
> Issue Type: Bug
> Reporter: Robert Stupp
> Attachments: 11818-direct-mem-unpooled.png, 11818-direct-mem.png,
> oom-histo-live.txt, oom-stack.txt
>
>
> The following stack trace is not caught by {{JVMStabilityInspector}}.
> Situation was caused by a load test with a lot of parallel writes and reads
> against a single node.
> {code}
> ERROR [SharedPool-Worker-1] 2016-05-17 18:38:44,187 Message.java:611 -
> Unexpected exception during request; channel = [id: 0x1e02351b,
> L:/127.0.0.1:9042 - R:/127.0.0.1:51087]
> java.lang.OutOfMemoryError: Direct buffer memory
> at java.nio.Bits.reserveMemory(Bits.java:693) ~[na:1.8.0_92]
> at java.nio.DirectByteBuffer.<init>(DirectByteBuffer.java:123)
> ~[na:1.8.0_92]
> at java.nio.ByteBuffer.allocateDirect(ByteBuffer.java:311)
> ~[na:1.8.0_92]
> at io.netty.buffer.PoolArena$DirectArena.newChunk(PoolArena.java:672)
> ~[netty-all-4.0.36.Final.jar:4.0.36.Final]
> at io.netty.buffer.PoolArena.allocateNormal(PoolArena.java:234)
> ~[netty-all-4.0.36.Final.jar:4.0.36.Final]
> at io.netty.buffer.PoolArena.allocate(PoolArena.java:218)
> ~[netty-all-4.0.36.Final.jar:4.0.36.Final]
> at io.netty.buffer.PoolArena.allocate(PoolArena.java:138)
> ~[netty-all-4.0.36.Final.jar:4.0.36.Final]
> at
> io.netty.buffer.PooledByteBufAllocator.newDirectBuffer(PooledByteBufAllocator.java:270)
> ~[netty-all-4.0.36.Final.jar:4.0.36.Final]
> at
> io.netty.buffer.AbstractByteBufAllocator.directBuffer(AbstractByteBufAllocator.java:177)
> ~[netty-all-4.0.36.Final.jar:4.0.36.Final]
> at
> io.netty.buffer.AbstractByteBufAllocator.directBuffer(AbstractByteBufAllocator.java:168)
> ~[netty-all-4.0.36.Final.jar:4.0.36.Final]
> at
> io.netty.buffer.AbstractByteBufAllocator.buffer(AbstractByteBufAllocator.java:105)
> ~[netty-all-4.0.36.Final.jar:4.0.36.Final]
> at
> org.apache.cassandra.transport.Message$ProtocolEncoder.encode(Message.java:349)
> ~[main/:na]
> at
> org.apache.cassandra.transport.Message$ProtocolEncoder.encode(Message.java:314)
> ~[main/:na]
> at
> io.netty.handler.codec.MessageToMessageEncoder.write(MessageToMessageEncoder.java:89)
> ~[netty-all-4.0.36.Final.jar:4.0.36.Final]
> at
> io.netty.channel.AbstractChannelHandlerContext.invokeWrite(AbstractChannelHandlerContext.java:619)
> ~[netty-all-4.0.36.Final.jar:4.0.36.Final]
> at
> io.netty.channel.AbstractChannelHandlerContext.write(AbstractChannelHandlerContext.java:676)
> ~[netty-all-4.0.36.Final.jar:4.0.36.Final]
> at
> io.netty.channel.AbstractChannelHandlerContext.write(AbstractChannelHandlerContext.java:612)
> ~[netty-all-4.0.36.Final.jar:4.0.36.Final]
> at
> org.apache.cassandra.transport.Message$Dispatcher$Flusher.run(Message.java:445)
> ~[main/:na]
> at
> io.netty.util.concurrent.PromiseTask$RunnableAdapter.call(PromiseTask.java:38)
> ~[netty-all-4.0.36.Final.jar:4.0.36.Final]
> at
> io.netty.util.concurrent.ScheduledFutureTask.run(ScheduledFutureTask.java:120)
> ~[netty-all-4.0.36.Final.jar:4.0.36.Final]
> at
> io.netty.util.concurrent.SingleThreadEventExecutor.runAllTasks(SingleThreadEventExecutor.java:358)
> ~[netty-all-4.0.36.Final.jar:4.0.36.Final]
> at io.netty.channel.nio.NioEventLoop.run(NioEventLoop.java:374)
> ~[netty-all-4.0.36.Final.jar:4.0.36.Final]
> at
> io.netty.util.concurrent.SingleThreadEventExecutor$2.run(SingleThreadEventExecutor.java:112)
> ~[netty-all-4.0.36.Final.jar:4.0.36.Final]
> at
> io.netty.util.concurrent.DefaultThreadFactory$DefaultRunnableDecorator.run(DefaultThreadFactory.java:137)
> ~[netty-all-4.0.36.Final.jar:4.0.36.Final]
> at java.lang.Thread.run(Thread.java:745) [na:1.8.0_92]
> {code}
> The situation does not get better when the load driver is stopped.
> I can reproduce this scenario at will. Managed to get histogram, stack traces
> and heap dump. Already increased {{-XX:MaxDirectMemorySize}} to {{2g}}.
> A {{nodetool flush}} causes the daemon to exit (as that direct-memory OOM is
> caught by {{JVMStabilityInspector}}).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
