Question

I am testing performance on integer addition in Java. The way I did that is by summing up billions of integers. The sample file I use for testing is a 1G binary file. My program is as simple as shown in the snippet below.

int result = 0;
FileChannel fileChannel = new FileInputStream(filename).getChannel();
long fileSize = fileChannel.size();
intBuffer = fileChannel.map(MapMode.READ_ONLY, startPosition, fileSize).asIntBuffer();

try {
  while (true) {
    result += intBuffer.get();
  }
} catch (BufferUnderflowException e) {
  System.out.println("Complete reading");
}

As you can see from above, it simply executes two operations in each loop

• read integer from file
• integer addition

This program ran about 2 minutes on my machine. I also did another test run without addition, by changing result += intBuffer.get() to result = intBuffer.get() (shown as in following snippet).

int result = 0;
FileChannel fileChannel = new FileInputStream(filename).getChannel();
long fileSize = fileChannel.size();
intBuffer = fileChannel.map(MapMode.READ_ONLY, startPosition, fileSize).asIntBuffer();

try {
  while (true) {
    result = intBuffer.get();
  }
} catch (BufferUnderflowException e) {
  System.out.println("Complete reading");
}

The entire program in this case turned out to complete within 1 second. Compared to its sibling variant above, it seems integer addition dominate the CPU time compared to IO read.

I wrote another benchmark program just for justify my guess, it does the same number of additions as the above example.

int result = random.nextInt();
int other = random.nextInt();
int num = 1073741824 / 4;
while(num-- > 0) {
  result += other;
}

With the same amount of integer additions plus the integer incremental operations, this program finishes less than 1 second.

My question is

• What caused the the major timing difference between these runs? Does Java compiler does something to optimize the last one?

Any thoughts are appreciated.

Answer 1

That’s because disk I/O is very slow compared the CPU.

In the first case, you’re reading from a file. So you’re bound by disk-access.

In the second case, it’s all in the CPU.

---

So this has nothing to do with the speed of addition.

• The first case is limited by the speed of your disk.
• The second case is (probably) limited by the speed of the random number generator.

---

As for why result = intBuffer.get() seems to be very fast: (pulled from comments)

Two possible reasons I can think of:

• Dead Code Elimination by the JIT is optimizing out all but the last iteration.
• I/O buffering: The OS is buffering the entire file into memory after the first read.*

*So subsequent passes will be very fast. It’s easy to test for this case by re-ordering the tests or clearing the I/O cache each time