Java app calls C++ DLL via JNI; how best to allocate memory?

Question

Basic summary of question is: How do I best optimize my memory allocation to give as much memory to the DLLs I access through JNI as possible? What should I aim to minimize, what should I aim to maximize, etc.

SYSTEM: Running JBoss 6 as a Windows 32 Service in a 32-bit system with 4 GB RAM. I do understand there are maximum restrictions on memory for Java Heap. JVM is JRE1.6.0_26

SERVICE: Installed under JBoss is a webapp which receives requests from clients; each request calls the C++-built DLL through JNI to process an image file in some fashion or other.

ISSUE: Occasionally, with larger or some (not all) LZW-compression images, the calling java class receives a message that the DLL experienced a Global Memory Depletion and failed to complete the requested process.

There is nothing else actively running on the server beyond basic windows processes.

Current JBOSS App Server memory settings are as follows, but may be excessive:

-Xms1024m -Xmx1024m -Xss1024k -XX:MaxPermSize=128m

I am trying to determine the best memory settings to give as much resources to the JNI DLL, as I understand JNI does not use any memory allocated to the Java Heap.

I have read these, but did not find them helpful to answer my question:

Java JNI : Memory allocation / partitioning

Can jconsole be used to identify memory leaks in JNI C++ objects?

The two answers currently supplied do not address the inherient question.

Current memory of JBoss server after one week with JVM params set as above (TaskManager indicates java.exe process at 750,672k)

Total Memory Pools: 5

Pool: Code Cache (Non-heap memory)

    Peak Usage : init:2359296, used:7317312, committed:7438336, max:50331648
    Current Usage : init:2359296, used:7306496, committed:7438336, max:50331648


        |---------| committed:7.09Mb
        +---------------------------------------------------------------------+
        |/////////| | max:48Mb
        +---------------------------------------------------------------------+
        |---------| used:6.97Mb


Pool: PS Eden Space (Heap memory)

    Peak Usage : init:268500992, used:354811904, committed:354811904, max:355991552
    Current Usage : init:268500992, used:270153472, committed:354091008, max:354156544


        |--------------------------------------------------------------------| committed:337.69Mb
        +---------------------------------------------------------------------+
        |///////////////////////////////////////////////////// || max:337.75Mb
        +---------------------------------------------------------------------+
        |----------------------------------------------------| used:257.64Mb


Pool: PS Survivor Space (Heap memory)

    Peak Usage : init:44695552, used:44694896, committed:78643200, max:78643200
    Current Usage : init:44695552, used:0, committed:1835008, max:1835008


        |---------------------------------------------------------------------| committed:1.75Mb
        +---------------------------------------------------------------------+
        | | max:1.75Mb
        +---------------------------------------------------------------------+
        | used:0b


Pool: PS Old Gen (Heap memory)

    Peak Usage : init:715849728, used:123671968, committed:715849728, max:715849728
    Current Usage : init:715849728, used:104048648, committed:715849728, max:715849728


        |---------------------------------------------------------------------| committed:682.69Mb
        +---------------------------------------------------------------------+
        |////////// | max:682.69Mb
        +---------------------------------------------------------------------+
        |---------| used:99.23Mb


Pool: PS Perm Gen (Non-heap memory)

    Peak Usage : init:16777216, used:91989664, committed:134217728, max:134217728
    Current Usage : init:16777216, used:90956472, committed:90963968, max:134217728


        |----------------------------------------------| committed:86.75Mb
        +---------------------------------------------------------------------+
        |//////////////////////////////////////////////| | max:128Mb
        +---------------------------------------------------------------------+
        |----------------------------------------------| used:86.74Mb

Answer 1

Memory allocated by the native code wrapped by JNI is allocated to the JVM process, but is not under the control of your Java code. It is not part of the heap, and is not tunable via JVM parameters. Basically, anything allocated with a native malloc must be managed by that native code. If you are in control of the libraries you are using, its imperative that you go through it and check for resource leaks. This is especially important if this is being used in a long lived process.

In my experience the best approach would be to examine your actual memory use by pulling the JMX stats exposed by the JVM. Once you have an idea about how much memory your Java app consumes You'll have a better idea about where to set your max heap settings. Permgen space is used for class definitions and such, so you really shouldn't need much memory there unless you are doing a bunch of dynamic class loading.

While you cannot tune the memory available for the JNI library, tuning the memory reserved for your heap and such will potentially free up resources for use by the library.

As would be expected, adding the heap memory peaks together it comes out to about 1022.19 (the max size of your heap). When the heap is exhausted a full GC run is kicked off and dirty heap is reclaimed. Based on the numbers that you have provided, I'd suggest starting with a Xmx512m. This will give your JNI code room to breath.

If you find that the JVM is thrashing due to excessive garbage collection, meaning that you're running out of Java heap too quickly, you could grow that allocation. However, if it is eating up 512mb rapidly enough to cause a noticeable performance impact, its unlikely that anything short of a significant increase will have much effect. This all depends heavily on your program, how quickly it eats the Java heap, and how effective the full GC run is.