Nio performance bottleneck

[JLM]

Hello !

In order to be sure that nio package was more performant than regular
streams, i have tested it with this little piece of code :

// ----------------------------
ByteBuffer loWriteBuffer = ByteBuffer.allocateDirect(1024);
CharBuffer loBuffer = CharBuffer.allocate(512);

Charset charset = Charset.forName("ISO-8859-1");
CharsetEncoder encoder = charset.newEncoder();
long llMillis, llMillis2, llMillis3;

llMillis = System.currentTimeMillis();
try {
FileChannel loChannel = new RandomAccessFile(new
File("C:\\Temp\\Test.txt"), "rw").getChannel();
for (int i = 0; i < 1000; ++i) {
loBuffer.put("ceci est un test\r\n");
loBuffer.flip();
encoder.encode(loBuffer, loWriteBuffer, false);
loWriteBuffer.flip();
loChannel.write(loWriteBuffer);
loBuffer.clear();
loWriteBuffer.clear();
}
loChannel.close();
}
catch (IOException leIO) {
leIO.printStackTrace();
}
llMillis2 = System.currentTimeMillis();
try {
Writer loWriter = new BufferedWriter(new OutputStreamWriter(new
FileOutputStream("C:\\Temp\\Test2.txt"), charset));
for (int i = 0; i < 1000; ++i) {
loWriter.write("ceci est un test\r\n");
}
loWriter.flush();
loWriter.close();
}
catch (IOException leIO) {
leIO.printStackTrace();
}
llMillis3 = System.currentTimeMillis();

System.out.println("Results : nio = " + (llMillis2 - llMillis) + " ms /
standard = " + (llMillis3 - llMillis2) + " ms");
// ----------------------------

I was surprised to see that, on my computer, the regular stream based part
was running about 4 times quicker than the nio based part !!
Here is my question : is my code correctlty optimized ? Anybody already
noticed this performance bottleneck ? Is my test the cause of this problem ?

Thank you !
Jean-Luc

 

[Yu SONG]

Hello !

In order to be sure that nio package was more performant than regular
streams, i have tested it with this little piece of code :
....

for (int i = 0; i < 1000; ++i) {
loBuffer.put("ceci est un test\r\n");
loBuffer.flip();
encoder.encode(loBuffer, loWriteBuffer, false);
loWriteBuffer.flip();
loChannel.write(loWriteBuffer);
loBuffer.clear();
loWriteBuffer.clear();
}
loChannel.close();
} ....

I was surprised to see that, on my computer, the regular stream based part
was running about 4 times quicker than the nio based part !!
Here is my question : is my code correctlty optimized ? Anybody already
noticed this performance bottleneck ? Is my test the cause of this problem ?

You don't need to encode and flip (twice) each time.

For most of my applications, I usually use a big enough buffer to
process all the data first, and then write this buffer to the file once
or twice.


--
Song

/* E-mail.c */
#define User "Yu.Song"
#define At '@'
#define Warwick "warwick.ac.uk"
int main() {
printf("Yu Song's E-mail: %s%c%s", User, At, Warwick);
return 0;}

Further Info. : http://www.dcs.warwick.ac.uk/~esubbn/
_______________________________________________________

 

[JLM]

You don't need to encode and flip (twice) each time.

For most of my applications, I usually use a big enough buffer to process
all the data first, and then write this buffer to the file once or twice.

It is only a test case, and it is intentionaly fractionned : i wanted to
test 1000 independent invocations of a complete write. You can assume that
each loop is not related with the others. But in fact you are right on a
point : my test is not good, because the two parts in my program have not
exactly the same behaviors. To really have two equivalent parts, i had to
change the source code on the twice part : i have to flush the output stream
in every loop, instead of flushing once at the end. In that case, the NIO
part is faster than the standard case. Honnor is safe ! ;-)
Thank you giving to me a clue to find the solution !