The following sections provide information on how to troubleshoot various problems that may arise for deployment in production environments.
Analyzing a stack trace
When your Java process starts to spin your CPU, you must immediately analyze the issue using the following two commands and obtain the invaluable information required to tackle the issue. This is done based on the process ID (pid).
jstack <pid> > thread-dump.txt
ps -C java -L -o pcpu,cpu,nice,state,cputime,pid,tid > thread-usage.txt
Tip: OS X users can alternatively use the command
ps M <PID>instead.
These commands provide you with the thread-dump.txt file and the thread-usage.txt file. After obtaining these two files, do the following.
Find the thread ID (the one that belongs to the corresponding PID) that takes up the highest CPU usage by examining the thread-usage.txt file.
In this example, the thread ID that takes up the highest CPU usage is 1604.
- Convert the decimal value (in this case 1604) to hexadecimal. You can use an online converter to do this. The hexadecimal value for 1604 is 644.
- Search the thread-dump.txt file for the hexadecimal obtained in order to identify the thread that spins. In this case, the hexadecimal value to search for is 644. The thread-dump.txt file should have that value as a thread ID of one thread.
That thread usually has a stack trace, and that's the lead you need to find the issue. In this example, the stack trace of the thread that spins is as follows.
Capturing the state of the system
Carbondump is a tool used to collect all the necessary data from a running WSO2 product instance at the time of an error. The carbondump generates a ZIP archive with the collected data that helps to analyze the system and determine the problem that caused the error. Therefore, it is recommended that you run this tool as soon as an error occurs in the WSO2 product instance.
When using the tool, you have to provide the process ID (pid) of the product instance and the
<PRODUCT_HOME> location, which is where your unzipped Carbon distribution files reside. The command takes the following format:
The tool captures the following information about the system:
- Operating system information** OS (kernel) version
- Installed modules lists and their information
- List of running tasks in the system
- Memory information of the Java process** Java heap memory dump
- Histogram of the heap
- Objects waiting for finalization
- Java heap summary. GC algo used, etc.
- Statistics on permgen space of Java heap
- Information about the running Carbon instance** Product name and version
- Carbon framework version (This includes the patched version)
- <PRODUCT_HOME>, <JAVA_HOME>
- configuration files
- log files
- H2 database files
- Thread dump
- Checksum values of all the files found in the $CARBON_HOME
Viewing process threads in Solaris
This information is useful to know in situations when the database processes are not fully utilizing the CPU's threading capabilities. It gives you a better understanding on how 11g and 10g takes advantage of threading and how you can validate those queries from the system.
The following information provides insight on whether a Solaris process is parallelized and is taking advantage of the threading within the CPU.
- Open a command line in Solaris.
prstatand have a look to the last column, labeled
PROCESS/NLWP. NLWP is a reference to the number of lightweight processes and are the number of threads the process is currently using with Solaris as there is a one-to-one mapping between lightweight processes and user threads. A single thread process will show
1there while a multi-threaded one will show a larger number. See the following code block for an example.
If you observe the
PROCESS/NLWPvalue in the example above, you can identify that
oracleare single thread processes, while
javais a multi-threaded process.
Alternatively, you can analyze individual thread activity of a multi-threaded process by using the
prstat -L -p pid. This displays a line for each thread sorted by CPU activity. In that case, the last column is labeled
PROCESS/LWPID, where LWPID is the thread ID. If more than one thread shows significant activity, your process is actively taking advantage of multi-threading.
Checking the health of a cluster
In Hazelcast, the health of a member in the cluster is determined by the heartbeats the member sends. If the well-known member does not receive a heartbeat within a given amount of time (this can be configured), then the node is assumed dead. By default, the given amount of time is 600 seconds (or 10mins), which might be too much for some scenarios.
Failure detectors used in distributed systems can be unreliable. In these sort of scenarios, Hazelcast uses heartbeat monitoring as a fault detection mechanism and the nodes send heartbeats to other nodes.
If a heartbeat message is not received by a given amount of time, Hazelcast assumes the node is dead. This is configured via the
hazelcast.max.no.heartbeat.seconds property. The optimum value for this property depends on the system. Although the default is 600 seconds, it might be necessary to reduce the heartbeat to a lower value if nodes are to be declared dead in a shorter time frame. However, you must verify this in your system and adjust as necessary depending on your scenario.
Warning: Reducing the value of this property to a lower value can result in nodes being considered as dead even if they are not. This results in multiple messages indicating that a node is leaving and rejoining the cluster.
Do the following steps to configure the maximum time between heartbeats.
- Create a property file called hazelcast.properties, and add the following property to it.
- Place this file in the
<PRODUCT_HOME>/repository/conf/directory in all the nodes in your cluster.
- Restart the servers.