NIO verses IO

Question

Hi All,
I am using java NIO socketchannel and ByteBuffers for my client/Server communication. It is in blocking mode and single threaded.
Problem Description: From the results it seems that my NIO program is taking more time than old java io. It looks strange as i was expecting
less time in communication using NIO.
In my old IO program i am using underlying socket streams.
Analysis results shows:
1. sun.nio.ch.SocketChannelImpl.write() taking more time than java.io.ObjectOutputStream.writeObject().
2. sun.nio.ch.SocketChannelImpl.read() taking more time than java.io.ObjectInputStream.readObject().
Comparison report between IO and NIO are (for sending/receiving 10,000 messages of around 9K size, msg size is not fixed):
Sender Program:
NIO Program:
Backtrace
Description of CPU usage for thread main
100.0% - 182635.0 ms - Sender_Final.main()
100.0% - 182635.0 ms - Sender_Final.query()
38.91% - 71078.0 ms - java.io.ObjectOutputStream.writeObject()
36.31% - 66321.0 ms - SerializedData.<init>()
*16.29% - 29769.0 ms - sun.nio.ch.SocketChannelImpl.write()*
6.63% - 12119.0 ms - java.io.ByteArrayOutputStream.toByteArray()
1.05% - 1935.0 ms - java.io.ObjectOutputStream.<init>()
0.08% - 158.0 ms - java.io.ByteArrayOutputStream.<init>()
0.07% - 138.0 ms - java.nio.DirectByteBuffer.put()
IO Program:
Backtrace
Description of CPU usage for thread main
100.0% - 23351.0 ms - SimpleSender.main()
100.0% - 23351.0 ms - SimpleSender.query()
*88.6% - 20690.0 ms - java.io.ObjectOutputStream.writeObject()*
9.96% - 2327.0 ms - SerializedData.<init>()
0.47% - 111.0 ms - java.io.ObjectOutputStream.reset()
0.05% - 12.0 ms - java.io.ObjectOutputStream.flush()
Receiver Programs:
NIO Program:
Backtrace
Description of CPU usage for thread main
100.0% - 176942.0 ms - Receiver_Final.main()
100.0% - 176942.0 ms - Receiver_Final.serve()
61.0% - 107935.0 ms - java.io.ObjectInputStream.readObject()
*34.75% - 61494.0 ms - sun.nio.ch.SocketChannelImpl.read()*
1.67% - 2957.0 ms - java.io.ObjectInputStream.<init>()
0.39% - 707.0 ms - java.nio.DirectByteBuffer.get()
0.09% - 163.0 ms - sun.nio.ch.ServerSocketChannelImpl.accept()
0.05% - 103.0 ms - java.io.ByteArrayInputStream.<init>()
IO Program:
Backtrace
Description of CPU usage for thread main
100.0% - 27854.0 ms - SimpleReceiver.main()
100.0% - 27854.0 ms - SimpleReceiver.serve()
*98.78% - 27515.0 ms - java.io.ObjectInputStream.readObject()*
0.13% - 37.0 ms - java.io.ObjectInputStream.<init>()
0.08% - 25.0 ms - java.net.ServerSocket.accept()
Any inputs here ? I am trying to achieve faster communication.
Does NIO always guarantee speed improvement over old IO? If not under what conditions it is advice to use.
I am using new InputStreams and OutputStreams for each message at both ends.
For sending data( Few code snippet):==>
while(true)
ByteArrayOutputStream bStream = new ByteArrayOutputStream();
ObjectOutputStream oStream = new ObjectOutputStream( bStream );
oStream.writeObject (serializedata);
byteArray = bStream.toByteArray();
objOutput.clear();
objOutput.put(byteArray);
objOutput.flip();
numWrite = sc.write(objOutput);
For receiving data ( Few code snippet):==>
while(offset < limit)
ByteArrayInputStream bStream = new ByteArrayInputStream(byteArray,offset ,limit - offset );
ObjectInputStream iStream = new ObjectInputStream( bStream);
serializeData = (SerializedData) iStream.readObject();
offset = limit-bStream.available();
}

Answer 1

Hi Sorabain,
I am not using ByteArrayOutputStream() in my old io programs and the code-snippet given here is for NIO program.
You can see in the performance comparion that old io is not using
ByteArrayOutputStream() or ByteArrayInputStream().
Yes for one thing i have to look into that why SerializedData<init> is taking more time in NIO in compare to old IO. As this is my serialized data class.
In my old IO program i am using underlying socket streams.
Please see a more detailed report below which has the comparion of socketwrite in old and new io both, probably it may throw more light on the cause:
Sender Program:
NIO Program:
Backtrace
Description of CPU usage for thread main
100.0% - 182635.0 ms - Sender_Final.main()
100.0% - 182635.0 ms - Sender_Final.query()
38.91% - 71078.0 ms - java.io.ObjectOutputStream.writeObject()
36.31% - 66321.0 ms - SerializedData.<init>()
*16.29% - 29769.0 ms - sun.nio.ch.SocketChannelImpl.write()*
6.63% - 12119.0 ms - java.io.ByteArrayOutputStream.toByteArray()
1.05% - 1935.0 ms - java.io.ObjectOutputStream.<init>()
0.08% - 158.0 ms - java.io.ByteArrayOutputStream.<init>()
0.07% - 138.0 ms - java.nio.DirectByteBuffer.put()
Hot spots
Name Percentage Time
SerializedData.<init> 36.31 66321.0 ms
java.io.ByteArrayOutputStream.write 34.38 62808.0 ms
sun.nio.ch.SocketDispatcher.write  15.38  28094.0 ms
java.io.ByteArrayOutputStream.toByteArray 6.63 12119.0 ms
java.io.ObjectOutputStream$BlockDataOutputStream.getUTFLength 1.14 2098.0 ms
java.io.ObjectOutputStream$BlockDataOutputStream.writeBytes 1.13 2076.0 ms
Sender_Final.query 0.61 1117.0 ms
java.nio.channels.spi.AbstractInterruptibleChannel.blockedOn 0.58 1076.0 ms
java.io.ObjectOutputStream.writeObject0 0.56 1040.0 ms
IO Program:
Backtrace
Description of CPU usage for thread main
100.0% - 23351.0 ms - SimpleSender.main()
100.0% - 23351.0 ms - SimpleSender.query()
88.6% - 20690.0 ms - java.io.ObjectOutputStream.writeObject()
9.96% - 2327.0 ms - SerializedData.<init>()
0.47% - 111.0 ms - java.io.ObjectOutputStream.reset()
0.05% - 12.0 ms - java.io.ObjectOutputStream.flush()
Hot spots
Name Percentage Time
java.net.SocketOutputStream.socketWrite  85.86  20051.0 ms
SerializedData.<init> 9.96 2327.0 ms
SimpleSender.query 0.9 211.0 ms
java.io.ObjectOutputStream$BlockDataOutputStream.writeBytes 0.8 188.0 ms
java.io.ObjectOutputStream.defaultWriteFields 0.6 141.0 ms
java.io.ObjectOutputStream$BlockDataOutputStream.writeInt 0.44 104.0 ms
java.io.ObjectOutputStream$BlockDataOutputStream.getUTFLength 0.42 99.0 ms
Receiver Programs:
NIO Program:
Backtrace
Description of CPU usage for thread main
100.0% - 176942.0 ms - Receiver_Final.main()
100.0% - 176942.0 ms - Receiver_Final.serve()
61.0% - 107935.0 ms - java.io.ObjectInputStream.readObject()
*34.75% - 61494.0 ms - sun.nio.ch.SocketChannelImpl.read()* 1.67% - 2957.0 ms - java.io.ObjectInputStream.<init>()
0.39% - 707.0 ms - java.nio.DirectByteBuffer.get()
0.09% - 163.0 ms - sun.nio.ch.ServerSocketChannelImpl.accept()
0.05% - 103.0 ms - java.io.ByteArrayInputStream.<init>()
Hot spots
Name Percentage Time
java.io.ObjectInputStream$BlockDataInputStream.readUTFSpan 42.62 75425.0 ms
sun.nio.ch.SocketDispatcher.read  34.27  60643.0 ms
java.io.ObjectInputStream$BlockDataInputStream.readUTFBody 10.77 19071.0 ms
Receiver_Final.serve 2.02 3583.0 ms
java.io.ObjectInputStream.readClassDescriptor 1.06 1891.0 ms
java.io.ObjectInputStream$BlockDataInputStream.<init> 0.85 1518.0 ms
java.io.ObjectInputStream$PeekInputStream.readFully 0.64 1139.0 ms
java.io.ByteArrayInputStream.read 0.63 1123.0 ms
java.io.ObjectStreamClass$FieldReflectorKey.<init> 0.52 933.0 ms
java.io.ObjectInputStream.readNonProxyDesc 0.48 855.0 ms
java.io.ObjectInputStream.readSerialData 0.44 781.0 ms
java.nio.DirectByteBuffer.get 0.39 707.0 ms
java.nio.channels.spi.AbstractInterruptibleChannel.blockedOn 0.39 702.0 ms
IO Program:
Backtrace
Description of CPU usage for thread main
100.0% - 27854.0 ms - SimpleReceiver.main()
100.0% - 27854.0 ms - SimpleReceiver.serve()
98.78% - 27515.0 ms - java.io.ObjectInputStream.readObject()
0.13% - 37.0 ms - java.io.ObjectInputStream.<init>()
0.08% - 25.0 ms - java.net.ServerSocket.accept()
Hot spots
Name Percentage Time
java.net.SocketInputStream.read  46.31  12900.0 ms
java.io.ObjectInputStream$BlockDataInputStream.readUTFBody 16.81 4683.0 ms
java.io.ObjectInputStream$BlockDataInputStream.readUTFSpan 14.5 4040.0 ms
java.net.PlainSocketImpl.acquireFD 1.64 459.0 ms
java.net.SocketInputStream.read 1.64 458.0 ms
java.net.PlainSocketImpl.releaseFD 1.55 432.0 ms
sun.security.provider.SHA.implCompress 1.09 305.0 ms
SimpleReceiver.serve 0.99 277.0 ms
java.io.ObjectInputStream.readString 0.95 266.0 ms
java.util.HashMap$EntrySet.iterator 0.94 264.0 ms
Edited by: herculeus on Oct 30, 2007 9:59 AM