I have 2 almost identical Dell boxes with properly updated Ubuntu 16.04 Unity. Let's call them Slow and Fast.
'free -h' output on Slow:
              total used free shared buff/cache available
Mem: 7,8G 1,3G 542M 8,8M 5,9G 6,1G
Swap: 2,9G 421M 2,5G

The same on Fast:
              total used free shared buff/cache available
Mem: 7,8G 1,0G 3,0G 23M 3,7G 6,4G
Swap: 14G 0B 14G

Both of them has dual-head accelerated cards, ATI Radeon on Slow and NVidia GeForce on Fast, and very similar monitors. Processors are i7-2600 and i7-3770, corresp.

Now, the problem: the excerpt from a test creating 1000 owned Windows always runs almost 8 minutes on Slow and 1m16s on Fast. This time doesn't depend much whether we use -mx128m flag or not.

import java.awt.Frame;
import java.awt.Window;
import java.lang.ref.WeakReference;
import java.util.Vector;

public class OwnedWindowsLeak
{
    public static void main(String[] args)
    {
        Frame owner = new Frame("F");

        // First, create many windows
        Vector<WeakReference<Window>> children =
            new Vector<WeakReference<Window>>();
        for (int i = 0; i < 1000; i++)
        {
            Window child = new Window(owner);
            children.add(new WeakReference<Window>(child));
        }
        owner.dispose();
    }
}

I'm making this bug P3 since the difference is so huge; please investigate, and if the reason allows, decrease the priority.