Memory notifications in Java

If you're like me, then you probably relied on the venerable Runtime.freeMemory() call to figure out how you application or container was doing. Especially if, (ahem) you might have suspected a memory leak (or two) at some point.

Well, JDK 1.5 has some fancy new stuff that's really nifty.

First off, you get a good picture of what your memory looks like (both heap and non-heap) like so:

    MemoryMXBean memBean = ManagementFactory.getMemoryMXBean();
    MemoryUsage heap = memBean.getHeapMemoryUsage();
    MemoryUsage nonHeap = memBean.getNonHeapMemoryUsage();
    System.out.println(heap);
    System.out.println(nonHeap);   

This gets you some fairly cool (and very bash-script analyzable) data:

init = 33161792(32384K) used = 301960(294K) 
  committed = 33226752(32448K) max = 512950272(500928K)
init = 19136512(18688K) used = 1913488(1868K) 
  committed = 19136512(18688K) max = 117440512(114688K)

Now, on to more interactive things:

public class Mem implements NotificationListener {
  
  public void handleNotification(Notification n, Object hb) {
    //-- we'll get to this in a bit
  }

  public static void main (String[] args) {
    Mem mem = new Mem();
    MemoryMXBean memBean = ManagementFactory.getMemoryMXBean();
    NotificationEmitter ne = (NotificationEmitter)memBean ;
    ne.addNotificationListener(mem, null, null);

    //-- more to come: configure thresholds
  }
}

The above class implements the javax.management.NotificationListener interface that defines the handleNotification method. Now, the main method fetches the MemoryMXBean which turns out to also be a NotificationEmitter (javadocs come in real handy). You could actually set up notifications where you define both filters and user objects can be applied and handed into the handleNotification call. We'll be amateurs and opt to filter nothing and not forward any of our objects.

With the notification mechanism in place, we need set up a threshold to be notified on. Expanding the main method:

  public static void main (String[] args) {
    Mem mem = new Mem();
    MemoryMXBean memBean = ManagementFactory.getMemoryMXBean();
    NotificationEmitter ne = (NotificationEmitter)memBean ;
    ne.addNotificationListener(mem, null, null);

    List memPools = ManagementFactory.getMemoryPoolMXBeans();
    for (Iterator i = memPools.iterator(); i.hasNext();) {
      MemoryPoolMXBean mp = i.next();
      if (mp.isUsageThresholdSupported() ) {
 // Found the heap! Let's add a notifier
 MemoryUsage mu = mp.getUsage();
 long max = mu.getMax();
 long alert = (max * 50)/100;
 System.out.println("Setting a warning on pool: " + mp.getName() + " for: " + alert);
 mp.setUsageThreshold(alert);
      }
    }

We can call for all the MemoryPoolMXBeans, and then check to see if threshold setting is supported. For those that we can configure, we set an alert for when we surpass 50% usage. At this point, you are in some pretty rarefied territory within the VM:

Setting a warning on pool: Code Cache for: 25165824
Setting a warning on pool: PS Old Gen for: 236748800
Setting a warning on pool: PS Perm Gen for: 33554432

So, can we avert the dreaded out of memory condition?

Here's our uber-malicious memory leak:

import java.util.*;

public class Leak extends Thread {

  public static boolean keepLeaking = true;

  public void run () {
    Map map = new HashMap();
    int x = 0;
    while (keepLeaking) {
      String key = new String(""+x*10000);
      String value = new String (""+x*243523);
      map.put(key, value);
      try { Thread.sleep(1); } catch (Exception e) {}
      x++;
    }
  }
}

A call to set keepLeaking to false can avert a sure VM memory issue. So, here's the fleshed out handleNotification method:

  public void handleNotification(Notification n, Object hb) {
    String type = n.getType();
    if (type.equals(MemoryNotificationInfo.MEMORY_THRESHOLD_EXCEEDED)) {
      // retrieve the memory notification information
      CompositeData cd = (CompositeData) n.getUserData();
      MemoryNotificationInfo memInfo = MemoryNotificationInfo.from(cd);
      System.out.println(memInfo.getPoolName() + " has exceeded the threshold : " +
        memInfo.getCount() + " times");
      System.out.println(memInfo.getUsage());
      Leak.keepLeaking = false;
    } else {
      System.out.println("Unknown notification: " + n);
    }
  }

And, presto:

$ java -Xmx8m Mem
Setting a warning on pool: Code Cache for: 25165824
Setting a warning on pool: PS Old Gen for: 3735552
Setting a warning on pool: PS Perm Gen for: 33554432
Leaks started
PS Old Gen has exceeded the threshold : 1 times
init = 1441792(1408K) used = 3763520(3675K) committed = 4849664(4736K) max = 7471104(7296K)
$

You'd probably want to reset the peak usage a few times and ensure that you see a consistent incursion before doing anything dramatic (give the GC a chance..). But, this is a fairly large leap from the prior polling mechanism that we would have had to adopt to monitor how a VM is doing. Now, we can code the reaction to a situation with set thresholds instead of managing polling threads.

List.remove

Here's a really simple one:

...
public static void main (String[] args) {
  List <String> l = new ArrayList<String>();
  String one = new String("one");
  System.out.println(l.size());
  l.add(one);
  l.add(one);
  System.out.println(l.size());
  l.remove(one);
  System.out.println(l.size());
}
...

That gets you this:

0
2
1

Well, it isn't rocket science. This isn't a Set, it is a List. And the method remove says:

Removes the first occurrence in this list of the specified element (optional operation). If this list does not contain the element, it is unchanged. More formally, removes the element with the lowest index i such that (o==null ? get(i)==null : o.equals(get(i))) (if such an element exists).

The block that does all the hard work in the ArrayList class looks like so:

            for (int index = 0; index < size; index++)
                if (o.equals(elementData[index])) {
                    fastRemove(index);
                    return true;
                }

So, if you want your output to look like so:

0
2
0

Change your list's remove to look like this:

   while (l.remove(one));

Or better yet, use a Set.

Cats and dog

The baby German Shepard babysat the cats for me last night.

Playing with NIO

I had heard a lot about how fast NIO was. So, I decided to find out. I picked up my BufferedReader benchmarking program and used it as a baseline. Then, I wrote the following to test how fast NIO ran:

....
      File f = new File("USA.txt");
      FileInputStream fin = new FileInputStream(f);
      FileChannel fc = fin.getChannel();
      long size = fc.size();
      ByteBuffer bb = ByteBuffer.allocate(2048);
      long start = System.currentTimeMillis();
      while ( (fc.read(bb)) > 0 ){
        bb.flip();
        String s = new String(bb.array());
      }
      long finish = System.currentTimeMillis();
      fc.close();
.....

With about a 100 runs, the two stack up pretty evenly to read in a file of about 3 MB.



Note that the first few runs see NIO slightly slower, and then it hits a step and speeds up again. Perhaps being a lot closer to the OS, you can actually see the contents of the file move through different cache layers as we hit it more.. Not sure.

This wasn't what I had expected. Perhaps the overhead on string creation was dwarfing the real gains in NIO. Or perhaps NIO performs better when the files to ingest are larger. So, I ran the benchmarks again with a file of about 200 MB.



NIO saves us 1 second over BufferedReader, which executes at approximately 5 seconds. Depending on whose side you are on, BufferedReader costs only 20% more, or NIO generates a savings of 25%. Regardless, NIO is faster. But, it is a little more complex too.

I don't know that the speed boost (specifically for File IO) warrants a switch to NIO. File locking sure seems nice, and it would be interesting to try and use NIO over a network to see if things change more dramatically there. Another point to keep in mind is that I used the allocate method in my NIO code, not the allocateDirect, which can optimize even further. However, in my scenario, the allocateDirect did not execute without a runtime exception.

To sum things up, NIO ran faster. The larger the file size, the greater the impact. For small files though, NIO could quite easily be overkill, both on the performance end, and the code maintenance end.

BufferedReader vs InputStream

Java has an IO system that's really flexible for reading and writing data. If you keep NIO out of the equation, then data from an input source typically originates from an InputStream. However, most tutorials and mentors would recommend that you wrap this raw stream into something more appropriate for your needs. With text files, when you want to inhale in Strings (XML falls into this category), you frequently see examples of an InputStream being wrapped up sequentially into a BufferedReader. So, do these abstraction layers for IO impose a penalty on performance?

I reused my big file (USA.txt) and set up a benchmark run with a BufferedReader.

...
     File f = new File("USA.txt");
     FileInputStream fin = new FileInputStream(f);
     BufferedReader in = new BufferedReader(new InputStreamReader(fin));
     String s = "";
     long start = System.currentTimeMillis();
     while ( (s = in.readLine()) != null) {
       int x = s.length();
     }
     long finish = System.currentTimeMillis();
     System.out.println(finish - start);
...

The chart shows the performance of 100 runs of the code. It's pretty stable and hovers somewhere between 80 to 100 milliseconds.



Next step: get the InputStream up and running and compare. So, I went with a code block like so:

...
      File f = new File("USA.txt");
      InputStream in = new FileInputStream(f);
      // buffSize configurable
      byte[] buffer = new byte[buffSize];
      int len = 0;
      long start = System.currentTimeMillis();
      while ( (len = in.read(buffer)) > 0 ) {
 String s = new String(buffer, 0, len);
 int x = s.length();
      }
      long finish = System.currentTimeMillis();
      System.out.println(finish - start);
...

One of the drawbacks to the InputStream is that you can either read a byte at a time, or have to supply a byte read size. I started low, with a 16 byte sized array to read into. I could have been cute here, but I opted for simplicity. Here's how the 16 byte sized reads from the InputStream stacked up against the BufferedReader:



Not so hot, but not so surprising either. With such a small buffer size, I was surely choking the read speeds. So, I decided to try the runs with incrementing buffer sizes (32, 64, 265, 512, 1024, 2048 and 2086). And, here's what happened:



As we ratchet up the buffer size, we see performance gains. However, these gains steadily diminish. The most notable detail is that there's no point at which the InputStream read outperforms the BufferedReader. Behind the nice and friendly methods that the BufferedReader provides, there's clearly a huge amount of cool code that bridges the Reader to the InputStream. The cute parts that figure out the best way to read seem to be hidden, and working.

Spend some time reading through performance myths (from the pros), and there's one constant that shows up consistently. Keep it simple. The VMs do a lot of hard work for you, so you can write code that is clean. And, clearly this simplicity doesn't impose a performance burden here. Nifty.

Now, I wonder how much NIO helps...

StringBuilder vs StringBuffer

As Java 1.5 came out, many of us were eager to get our hands on StringBuilder. As a non-thread-safe version of StringBuffer, one would imagine that it would pack some heat on the performance end for cases where you don't have multi-threaded appends.

To test the relative performance differences, I used a BufferedReader over decently sized text file (622732 words gathered by repeatedly pasting the wikipedia document on USA). Armed with data, I wrote and measured the following loops that merely read in the file and appended to either a StringBuffer or a StringBuilder.

....
    for (int x=0; x<100; x++) {
      File f = new File("USA.txt");
      BufferedReader in = new BufferedReader(new InputStreamReader(new FileInputStream(f)));
      String line = "";
      StringBuffer sb = new StringBuffer(); 
      // StringBuilder sb = new StringBuilder();
      long start = System.currentTimeMillis();
      while ((line = in.readLine()) != null) {
        sb.append(line);
      }
      long mid = System.currentTimeMillis();
      String s = sb.toString();
      long done = System.currentTimeMillis();
      System.out.println((mid - start) + " " + (done - mid));
      in.close();
    }
...

As evident, I merely switched the constructor to a StringBuilder when I was done. The measurement dumps out the total file read and append times as well as to hit on the toString method.



Clearly, they both perform on almost the same lines. At Java One last year, there was lots of talk about how well the Sun VM was handling synchronization. This is proof.

But here's the dicey part. When we did a mass search and replace on our application, we found a massive boost. So, were we dreaming?

I was convinced that my findings were bogus, and that the overhead imposed by the readLine call on my measurement was hiding a performance difference. I was wrong. I changed the loop to measure like so:

...
      long read = 0;
      while ((line = in.readLine()) != null) {
        long t1 = System.currentTimeMillis();
        sb.append(line);
        long t2 = System.currentTimeMillis();
        read += (t2 - t1);
      }
...

No difference. I'm happy about this because we need not start hunting down StringBuffer and switching it to StringBuilder like crazy. But, I'm perplexed about what we saw earlier. So, I decided to try and execute the same test on a JRockit VM (1.5 spec). Here's what I got:



JRockit seems to jump in performance steps- almost like the VM is adjusting to the code. Notice the step like decrease in append times for both the buffer and the builder, and the toString is clearly faster than Sun. But, there's something to be said for the sheer predictability of the Sun VM too.

So, my conclusion? Don't race to switch StringBuffer to StringBuilder- there doesn't seem to be a real tangible difference in performance.

entry zero

In looking up the term recursion on wikipedia, I found a nifty item: The Droste Effect. MC Escher appears to have made use of the technique.