Java 7, ulimit, and non-heap memory

A couple of co-workers of mine found the following problem:

Error occurred during initialization of VM
  Could not reserve enough space for object heap
  Could not create the Java virtual machine.

It happened in a batch-job environment where we use ulimit (for all programs, Java or not), in addition to Java's heap specification with -Xms and -Xmx. And it only started happening on a test switch from Java 6 to Java 7.

Now, Java requires memory for more than just heap --- native C malloc in JNI, mmap, and so on. But it also (moreso in Java 7, which is the point of this post) needs some non-heap memory to keep track of the heap itself.

A way to measure this is the following script called trymem:

#!/bin/bash
if [ $# -ne 3 ]; then
    echo "Usage: $0 {javadir} {heap MB} {extra MB}"
    exit 1
fi
javadir=$1
heap_mb=$2
xtra_mb=$3

totl_mb=$[heap_mb+xtra_mb]
totl_kb=$[totl_mb*1024]

ulimit -v $totl_kb

$javadir/bin/java -cp . -Xmx${heap_mb}m -Xms${heap_mb}m MyProgram
status=$?
echo ${heap_mb}+${xtra_mb}:${status}
exit $status

where MyProgram.java is simply

public class MyProgram {
    public static void main(String[] args) {
    }
}

along with a second script called searchmem:

#!/bin/bash

if [ $# -ne 1 ]; then
    echo "Usage: $0 {Java dir}"
    exit 1
fi
javadir=$1

heap_mb=10000
while [ $heap_mb -le 60000 ]; do
    echo -n "$heap_mb "
    xtra_mb=100
    while true; do
        ./trymem $javadir $heap_mb $xtra_mb 1> /dev/null 2> /dev/null
        status=$?
        echo -n .
        if [ $status -eq 0 ]; then
            echo $xtra_mb
            break
        elif [ $xtra_mb -gt $heap_mb ]; then
            echo "> $heap_mb"
            break
        fi
        xtra_mb=$[xtra_mb+100]
    done
    heap_mb=$[heap_mb+5000]
done

The idea is to set ulimit to the heap size, then keep increasing it until the test program can run without error. The difference between Java 6 and Java 7 is significant:

$ ./searchmem /usr/local/jdk/x86_64/jdk1.6.0_35

10000 ....400
15000 ....400
20000 ....400
25000 ....400
30000 .....500
35000 .....500
40000 .....500
45000 .....500
50000 .....500
55000 ......600
60000 ......600

$ ./searchmem /usr/local/jdk/x86_64/jdk1.7.0_9

10000 ........800
15000 ..........1000
20000 .............1300
25000 ..............1400
30000 ................1600
35000 ...................1900
40000 .....................2100
45000 .......................2300
50000 .........................2500
55000 ............................2800
60000 .............................2900

We couldn't find a pronouncement from Oracle regarding minimum ulimit as a function of heap size. But a plot of the above numbers suggests a linear relationship, and a regression (with more densely sampled data in searchmem, stepping heap_mb by 1000 rather than 5000) shows slope 1/24 and intercept 400MB. That is, given a Java heap size in MB, divide it by 24 and add 400 to it to find Java 7's heap-management overhead.

P.S. Thanks to David Craft for syntax highlighting in Blogspot:
http://www.craftyfella.com/2010/01/syntax-highlighting-with-blogger-engine.html