Memory and Threading

Because a single PXF Service (JVM) serves multiple segments on a segment host, the PXF heap size can be a limiting runtime factor. This becomes more evident under concurrent workloads or with queries against large files. You may run into situations where a query hangs or fails due to insufficient memory or the Java garbage collector impacting response times. To avert or remedy these situations, first try increasing the Java maximum heap size or decreasing the Tomcat maximum number of threads, depending upon what works best for your system configuration. You may also choose to configure PXF to perform specific actions when it detects an out of memory condition.

Increasing the JVM Memory for PXF

Each PXF Service running on a Greenplum Database host is configured with a default maximum Java heap size of 2GB and an initial heap size of 1GB. If the hosts in your Greenplum Database cluster have an ample amount of memory, try increasing the maximum heap size to a value between 3-4GB. Set the initial and maximum heap size to the same value if possible.

Perform the following procedure to increase the heap size for the PXF Service running on each host in your Greenplum Database cluster.

  1. Log in to your Greenplum Database master node:

    $ ssh gpadmin@<gpmaster>
    
  2. Edit the $PXF_BASE/conf/pxf-env.sh file. For example:

    gpadmin@gpmaster$ vi $PXF_BASE/conf/pxf-env.sh
    
  3. Locate the PXF_JVM_OPTS setting in the pxf-env.sh file, and update the -Xmx and/or -Xms options to the desired value. For example:

    PXF_JVM_OPTS="-Xmx3g -Xms3g"
    
  4. Save the file and exit the editor.

  5. Use the pxf cluster sync command to copy the updated pxf-env.sh file to the Greenplum Database cluster. For example:

    gpadmin@gpmaster$ pxf cluster sync
    
  6. Restart PXF on each Greenplum Database host as described in Restarting PXF.

Configuring Out of Memory Condition Actions

In an out of memory (OOM) situation, PXF returns the following error in response to a query:

java.lang.OutOfMemoryError: Java heap space

You can configure the PXF JVM to enable/disable the following actions when it detects an OOM condition:

  • Auto-kill the PXF Service (enabled by default).
  • Dump the Java heap (disabled by default).

Auto-Killing the PXF Server

By default, PXF is configured such that when the PXF JVM detects an out of memory condition on a Greenplum host, it automatically runs a script that kills the PXF Service running on the host. The PXF_OOM_KILL environment variable in the $PXF_BASE/conf/pxf-env.sh configuration file governs this auto-kill behavior.

When auto-kill is enabled and the PXF JVM detects an OOM condition and kills the PXF Service on the host:

  • PXF logs the following messages to $PXF_LOGDIR/pxf-oom.log on the segment host:

    =====> <date> PXF Out of memory detected <======
    =====> <date> PXF shutdown scheduled <======
    =====> <date> Stopping PXF <======
    
  • Any query that you run on a PXF external table will fail with the following error until you restart the PXF Service on the host:

    ... Failed to connect to <host> port 5888: Connection refused
    

When the PXF Service on a host is shut down in this manner, you must explicitly restart the PXF Service on the host. See the pxf reference page for more information on the pxf start command.

Refer to the configuration procedure below for the instructions to disable/enable this PXF configuration property.

Dumping the Java Heap

In an out of memory situation, it may be useful to capture the Java heap dump to help determine what factors contributed to the resource exhaustion. You can configure PXF to write the heap dump to a file when it detects an OOM condition by setting the PXF_OOM_DUMP_PATH environment variable in the $PXF_BASE/conf/pxf-env.sh configuration file. By default, PXF does not dump the Java heap on OOM.

If you choose to enable the heap dump on OOM, you must set PXF_OOM_DUMP_PATH to the absolute path to a file or directory:

  • If you specify a directory, the PXF JVM writes the heap dump to the file <directory>/java_pid<pid>.hprof, where <pid> identifies the process ID of the PXF Service instance. The PXF JVM writes a new file to the directory every time the JVM goes OOM.
  • If you specify a file and the file does not exist, the PXF JVM writes the heap dump to the file when it detects an OOM. If the file already exists, the JVM will not dump the heap.

Ensure that the gpadmin user has write access to the dump file or directory.

Note: Heap dump files are often rather large. If you enable heap dump on OOM for PXF and specify a directory for PXF_OOM_DUMP_PATH, multiple OOMs will generate multiple files in the directory and could potentially consume a large amount of disk space. If you specify a file for PXF_OOM_DUMP_PATH, disk usage is constant when the file name does not change. You must rename the dump file or configure a different PXF_OOM_DUMP_PATH to generate subsequent heap dumps.

Refer to the configuration procedure below for the instructions to enable/disable this PXF configuration property.

Procedure

Auto-kill of the PXF Service on OOM is enabled by default. Heap dump generation on OOM is disabled by default. To configure one or both of these properties, perform the following procedure:

  1. Log in to your Greenplum Database master node:

    $ ssh gpadmin@<gpmaster>
    
  2. Edit the $PXF_BASE/conf/pxf-env.sh file. For example:

    gpadmin@gpmaster$ vi $PXF_BASE/conf/pxf-env.sh
    
  3. If you want to configure (i.e. turn off, or turn back on) auto-kill of the PXF Service on OOM, locate the PXF_OOM_KILL property in the pxf-env.sh file. If the setting is commented out, uncomment it, and then update the value. For example, to turn off this behavior, set the value to false:

    export PXF_OOM_KILL=false
    
  4. If you want to configure (i.e. turn on, or turn back off) automatic heap dumping when the PXF Service hits an OOM condition, locate the PXF_OOM_DUMP_PATH setting in the pxf-env.sh file.

    1. To turn this behavior on, set the PXF_OOM_DUMP_PATH property value to the file system location to which you want the PXF JVM to dump the Java heap. For example, to dump to a file named /home/gpadmin/pxfoom_segh1:

      export PXF_OOM_DUMP_PATH=/home/pxfoom_segh1
      
    2. To turn off heap dumping after you have turned it on, comment out the PXF_OOM_DUMP_PATH property setting:

      #export PXF_OOM_DUMP_PATH=/home/pxfoom_segh1
      
  5. Save the pxf-env.sh file and exit the editor.

  6. Use the pxf cluster sync command to copy the updated pxf-env.sh file to the Greenplum Database cluster. For example:

    gpadmin@gpmaster$ pxf cluster sync
    
  7. Restart PXF on each Greenplum Database host as described in Restarting PXF.

Another Option for Resource-Constrained PXF Segment Hosts

If increasing the maximum heap size is not suitable for your Greenplum Database deployment, try decreasing the number of concurrent working threads configured for PXF’s embedded Tomcat web server. A decrease in the number of running threads will prevent any PXF node from exhausting its memory, while ensuring that current queries run to completion (albeit a bit slower). Tomcat’s default behavior is to queue requests until a thread is free, or the queue is exhausted.

The default maximum number of Tomcat threads for PXF is 200. The pxf.max.threads property in the pxf-application.properties configuration file controls this setting.

If you plan to run large workloads on a large number of files in an external Hive data store, or you are reading compressed ORC or Parquet data, consider specifying a lower pxf.max.threads value. Large workloads require more memory, and a lower thread count limits concurrency, and hence, memory consumption.

Note: Keep in mind that an increase in the thread count correlates with an increase in memory consumption.

Perform the following procedure to set the maximum number of Tomcat threads for the PXF Service running on each host in your Greenplum Database deployment.

  1. Log in to your Greenplum Database master node:

    $ ssh gpadmin@<gpmaster>
    
  2. Edit the $PXF_BASE/conf/pxf-application.properties file. For example:

    gpadmin@gpmaster$ vi $PXF_BASE/conf/pxf-application.properties
    
  3. Locate the pxf.max.threads setting in the pxf-application.properties file. If the setting is commented out, uncomment it, and then update to the desired value. For example, to reduce the maximum number of Tomcat threads to 100:

    pxf.max.threads=100
    
  4. Save the file and exit the editor.

  5. Use the pxf cluster sync command to copy the updated pxf-application.properties file to the Greenplum Database cluster. For example:

    gpadmin@gpmaster$ pxf cluster sync
    
  6. Restart PXF on each Greenplum Database host as described in Restarting PXF.