Are the ByteBuffer/IntBuffer/ShortBuffer Java classes fast?

Question

I'm working on an Android application (in Java, obviously) and I recently updated my UDP reader code. In both versions, I set up some buffers and receive a UDP packet:

byte[] buf = new byte[10000];
short[] soundData = new short[1000];
DatagramPacket packet = new DatagramPacket (buf, buf.length);
socket.receive (packet);

In the initial version, I put the data back together one byte at a time (it's actually 16 PCM audio data):

for (int i = 0; i < count; i++)
    soundData[i] = (short) (((buf[k++]&0xff) << 8) + (buf[k++]&0xff));

In the updated version, I used some cool Java tools I didn't know about when I started:

bBuffer  = ByteBuffer.wrap (buf);
sBuffer  = bBuffer.asShortBuffer();
sBuffer.get (soundData, 0, count);

In both cases, "count" is being populated correctly (I checked). However, there appear to be new problems with my streaming audio -- perhaps it isn't being handled fast enough -- which doesn't make any sense to me. Obviously, the buffer code is compiling into a lot more than three statements of JVM code, but it sure seemed like a reasonable assumption when I start this that the 2nd version would be faster than the 1st.

Patently, I'm not insisting that my code HAS to use Java NIO buffers, but at first glance at least, it DOES seem like a mo' betta' to go about this.

Anybody got any recommendations for a fast, simple Java UDP reader and whether there is a generally accepted "best way"??

Thanks, R.

Answer 1

Your code would be more efficient if instead of reading a packet into a byte array (copying the data from a native buffer into the array) and then wrapping it in a new ByteBuffer (creating a new object) and converting to a ShortBuffer (creating a new object) you set up your objects only once and avoided the copy.

You can do this by using DatagramChannel.socket() to create your socket, then connecting it as usual and usuing socket.getChannel() to get a DatagramChannel object. This object will allow you to read packets directly into an existing ByteBuffer (which you should create with ByteBuffer.allocateDirect for maximum efficiency). You can then us asShortBuffer() just once to create a view of your data as shorts, and read from that ShortBuffer after every time you refill the ByteBuffer.

The code therefore looks like this:

 DatagramSocket socket = DatagramChannel.socket();
 // code to connect socket
 DatagramChannel channel = socket.getChannel();
 ByteBuffer buffer = ByteBuffer.allocateDirect (10000);
 // you may want to invoke buffer.order(...) here to tell it what byte order to use
 ShortBuffer shortBuf = buffer.asShortBuffer();

 // in your receive loop:
 buffer.clear();
 channel.receive(buffer);
 shortBuf.position(0).limit(buffer.position()/2); // may ignore a byte if odd number received
 shortBuf.get(soundBuf,0,shortBuf.limit());

You should find this is much more efficient than your previous code because it avoids an entire copy of the data and the format conversion is handled by hand-optimized code rather than compiler generated byte manipulation which may be suboptimal. It will be somewhat more efficient if you use the platform-native byte order (I believe Android uses little-endian byte order on all platforms it is available for, and your code above seems to be big-endian, so this may not be possible for you), in which case shortBuf.get() becomes a direct memory copy.

Answer 2

In general, working with primitive types directly is going to be more efficient than working with objects because you avoid some of the overhead of creating objects, function calls, etc.

There are reasons to use the utility objects other than speed: convenience, safety, etc.

The best way to test the difference in this particular case would be to actually measure it. Try out both methods with a large dataset and time it. Then, you can decide if it is worth the benefits in this case.

You can also use Android's profiler to see where your problems really are. See TraceView.

Answer 3

I would use DataInputStream for this task, wrapped around a ByteArrayInputStream wrapped around the byte array. Encapsulates the shifting in readShort() without the overheads of ByteBuffer.

Alternatively you could read directly into a DirectByteBuffer via a DatagramChannel, rather than using DatagramPacket & DatagramSocket. At present you're using a mixture of java.net and java.nio.

I wouldn't expect major performance differences between these two approaches, but I would expect them both to be faster than your hybrid approach.