Broker (5.5) Deadlock when limiting size of temp store, and using VirtualTopics
-------------------------------------------------------------------------------
Key: AMQ-3530
URL: https://issues.apache.org/jira/browse/AMQ-3530
Project: ActiveMQ
Issue Type: Bug
Components: Broker
Affects Versions: 5.5.0
Environment: 10.8.0 Darwin Kernel Version 10.8.0: Tue Jun 7 16:33:36
PDT 2011; root:xnu-1504.15.3~1/RELEASE_I386 i386
Reporter: Dominic Tootell
When using Virtual Topics and attempting to limit the Temp Table space the
broker will dead lock.
I have created a test case the demonstrates the above issue. The test case has
the following scenario:
- 1 Producer Thread (own connection) writing to VirtualTopic.FooTwo
-- writes 10000 messages
- 2 Consumer Threads (each own connection) consuming from
Consumer.X.VirtualTopic.FooTwo
-- consumes message (auto ack - but message ack for good measure - delay of
10ms between consumption)
The dead lock occurs when using the either the KahaDB or ActiveMQPersistence
persistence store is used, and seems more related to the KahaDB implementation
used for the Temp storage area.
I shall attach the test case (maven project), which when built (mvn clean
package -DskipTests=true) will create an executable jar
(target/5.5.0-deadlock-jar-with-dependencies.jar), from which the issue can be
replicated:
The tests can be run as follows:
{noformat}
java -classpath target/5.5.0-deadlock-jar-with-dependencies.jar
bbc.forge.domt.activemq.investigation.KahaDBTempStorageDeadlockReplication55
{noformat}
or
{noformat}
java -classpath target/5.5.0-deadlock-jar-with-dependencies.jar
bbc.forge.domt.activemq.investigation.TempStorageDeadlockReplication55
{noformat}
These classes are also Unit Test Cases, and output the following log files:
- Producer logs to target/producer.log
- Consumes log to target/consumer.log
- A Monitor thread just run in the background that detail the number of
messages sent and consumed... logs to target/monitor.log
- tests write to the dir *{{target/activemq-data}}*
The target/monitor.log can be tailed, upon which you can see the dead lock
occur; where the consumer stops consumering and the producer stops sending;
something like the following:
{noformat}
Dominic-Tootells-MacBook-Pro-2:trunk dominict$ tail -f monitor.log
[2011.10.08 17:15:09] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Sent(Produced):0
[2011.10.08 17:15:09] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Consumed:0
[2011.10.08 17:15:10] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Sent(Produced):248
[2011.10.08 17:15:10] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Consumed:35
[2011.10.08 17:15:11] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Sent(Produced):248
[2011.10.08 17:15:11] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Consumed:74
[2011.10.08 17:15:12] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Sent(Produced):248
[2011.10.08 17:15:12] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Consumed:114
[2011.10.08 17:15:13] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Sent(Produced):248
[2011.10.08 17:15:13] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Consumed:152
[2011.10.08 17:15:14] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Sent(Produced):248
[2011.10.08 17:15:14] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Consumed:192
[2011.10.08 17:15:15] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Sent(Produced):248
[2011.10.08 17:15:15] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Consumed:232
[2011.10.08 17:15:16] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Sent(Produced):248
[2011.10.08 17:15:16] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Consumed:248
[2011.10.08 17:15:17] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Sent(Produced):248
[2011.10.08 17:15:17] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Consumed:248
[2011.10.08 17:15:18] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Sent(Produced):248
[2011.10.08 17:15:18] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Consumed:248
[2011.10.08 17:15:19] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Sent(Produced):248
[2011.10.08 17:15:19] [Thread-14] INFO MonitorConsumerAndProducedThread - No
of Message Consumed:248
{noformat}
To disable limiting the temp storage add the System property
*{{limit.temp=false}}*; and you see that the deadlock no longer occurs:
{noformat}
java -Dlimit.temp=false -classpath
target/5.5.0-deadlock-jar-with-dependencies.jar
bbc.forge.domt.activemq.investigation.KahaDBTempStorageDeadlockReplication55
{noformat}
----
The tests can be run as maven tests directly:
{noformat}
mvn clean test -Dtest=KahaDBTempStorageDeadlockReplication55Test
-Dlimit.temp=true
{noformat}
----
This seems to be a different/additional issue to AMQ-2475
----
The reason why this issue affects our ActiveMQ environment/installation so
much, is that we use a shared infrastructure model. Meaning that we provision
many
activemq instances on the same physcial hardware for different
projects/clients. We limit the disk space assigned to each broker so that no
one application goes wild and consumes all the disk space on the hardware; and
as a result impacts other projects/clients. Although it theory
running more than one instance of ActiveMQ on a server might seem counter
intutitive (with regards to performance - disk seeking); this is a
business model by which we can get the most money out of out physical hardware
and provide a MOM for most projects; for which maximum throughput is a
secondary concern.
Apologies that this issue is a duplicate of AMQ-3061. However, the reason I'm
opening an additional ticket is the hope that I can actually provide you a
patch in here (if possible)
that is coded against the 5.5 apache activemq version. As in AMQ-3061 I
accidentally:
a) provided a faulty patch
b) gave you a patch against a 5.4.1 version of fuse's activemq.
Big appologies for that.
thanks in advance, let me know if you need any more information.
/dom
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
