TimingTest: C# vs Java - Computer Technical Help

Two Plus Two Forums Other Topics Computer and Technical Help

TimingTest: C# vs Java

Post Reply Subscribe

...

Page 1 of 2

1 2

Page 1 of 2

1 2

04-21-2013 , 08:37 PM

:::grimReaper:::

veteran

Join Date: Jul 2010 Posts: 3,113

Testing how fast C# and Java are able to compute sqrt, log, exp etc on a billion iterations. Something seems wrong, because C# is lagging by a lot. Anything I'm doing wrong? Fwiw, I'm running the C# .exe from bin/release.

Java (~320ms):

Code:

import java.text.DecimalFormat;

public class TimingTest {
	public static void main(String[] args){
		final long maxIterations = 1000000000;
		DecimalFormat formatter = new DecimalFormat("0.###E0");
		String f = formatter.format(maxIterations);
		System.out.println("Beginning test on " + f + " iterations...");
		final long start = System.currentTimeMillis();
		for (long i = 1; i < maxIterations; i++) Math.sqrt(i);
		final long stop = System.currentTimeMillis();
		final long elapsed = (stop - start); 
		System.out.println("Time elapsed: " + elapsed);
	}
}

C# (~950ms):

Code:

namespace TimingTest
{
    class Program
    {
        static void Main(string[] args)
        {
            long maxIterations = 1000000000;
            Console.WriteLine("Beginning test on " + maxIterations.ToString("0.###E0") + " iterations...");
            DateTime start = DateTime.Now;
            for (long i = 1; i < maxIterations; i++) Math.Sqrt(i);
            double elapsed = DateTime.Now.Subtract(start).TotalMilliseconds;
            Console.WriteLine("Time elapsed: " + elapsed);
            Console.Read();
        }
    }
}

Python (too slow...):

Code:

from math import sqrt
maxIterations = 1000000000
i = 1

while i < maxIterations:
    sqrt(i)
    i = i + 1

Quote

150% up to $2,000 Welcome Bonus on CoinPoker

Join the action now

Daily Rewards ? Splash Pots ? CoinRaces

04-21-2013 , 11:45 PM

spiral

grinder

Join Date: Jan 2003 Posts: 528

Code:

from math import sqrt
maxIterations = 1000000000
i = 1

while i < maxIterations:
    sqrt(i)
    i = i + 1

Complete tangent but "for i in xrange(maxIterations):" should be 30% faster or so than this while loop

Quote

04-22-2013 , 01:01 AM

sards

old hand

Join Date: Dec 2006 Posts: 1,554

It seems unlikely that your CPU can do a billion square roots in 350ms. So the compiler is probably optimizing it out. Maybe try something like this:

Code:

...
double total = 0.0;
for (long i = 0; i < maxIterations; i++)
    total += Math.sqrt(i);
System.out.println("Total: " + total);
...

Quote

04-22-2013 , 01:18 AM

jmark

Pooh-Bah

Join Date: Dec 2003 Posts: 3,790

Also what optimization level are you using?

To avoid the compiler doing it for you you could populate an array with random numbers first and then take the square root of each.

Quote

04-22-2013 , 01:29 AM

:::grimReaper:::

veteran

Join Date: Jul 2010 Posts: 3,113

Quote:

Originally Posted by spiral

Complete tangent but "for i in xrange(maxIterations):" should be 30% faster or so than this while loop

Thanks, I used range() initially and got a memory problem. Clearly still learning Python. But in any case (xrange or for), the program is so slow it didn't terminate even after 2 minutes and not accumulating the sum like below.

Quote:

Originally Posted by sards

It seems unlikely that your CPU can do a billion square roots in 350ms. So the compiler is probably optimizing it out. Maybe try something like this:

Java: ~8400ms
C#: ~17000ms

Quote

04-22-2013 , 02:34 AM

:::grimReaper:::

veteran

Join Date: Jul 2010 Posts: 3,113

Quote:

Originally Posted by jmark

Also what optimization level are you using?

I don't know what you mean.

Quote:

Originally Posted by jmark

To avoid the compiler doing it for you you could populate an array with random numbers first and then take the square root of each.

This was it. I don't know much about compilers. Can anyone explain what's going on here? And in general, can I expect much performance difference between Java and C#?

On 100 million random square roots on numbers between 1 and 100:
Java: 452
C#: 442

Java

Code:

import java.text.DecimalFormat;

public class TimingTest {
	public static void main(String[] args){
		final int maxIterations = 100000000;
		double total = 0.0;
		
		long[] rnd = new long[maxIterations];
		for (int i = 0; i < maxIterations; i++) 
			rnd[i] = Math.round(Math.random()*100);
		
		DecimalFormat formatter = new DecimalFormat("0.###E0");
		String f = formatter.format(maxIterations);
		System.out.println("Beginning test on " + f + " iterations...");
		
		final long start = System.currentTimeMillis();
		for (int i = 0; i < maxIterations; i++) total += Math.sqrt(rnd[i]);
		final long stop = System.currentTimeMillis();
		
		final long elapsed = (stop - start); 
		System.out.println("Time elapsed: " + elapsed);
		System.out.println("Sum: " + total);
	}
}

Code:

using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Threading.Tasks;

namespace TimingTest
{
    class Program
    {
        static void Main(string[] args)
        {
            int maxIterations = 100000000;
            double total = 0.0;

            int[] rnd = new int[maxIterations];
            Random rng = new Random();
            for (int i = 0; i < maxIterations; i++) 
                rnd[i] = rng.Next(100);

            Console.WriteLine("Beginning test on " + maxIterations.ToString("0.###E0") + " iterations...");

            DateTime start = DateTime.Now;
            for (long i = 0; i < maxIterations; i++) total += Math.Sqrt(rnd[i]);
            double elapsed = DateTime.Now.Subtract(start).TotalMilliseconds;

            Console.WriteLine("Time elapsed: " + elapsed);
            Console.WriteLine("Sum: " + total);
            Console.Read();
        }
    }
}

Quote

04-22-2013 , 08:53 AM

adios

Carpal \'Tunnel

Join Date: Sep 2002 Posts: 22,558

Quote:

Originally Posted by :::grimReaper:::

I don't know what you mean.

This was it. I don't know much about compilers. Can anyone explain what's going on here? And in general, can I expect much performance difference between Java and C#?

On 100 million random square roots on numbers between 1 and 100:
Java: 452
C#: 442

Java

Code:

import java.text.DecimalFormat;

public class TimingTest {
	public static void main(String[] args){
		final int maxIterations = 100000000;
		double total = 0.0;
		
		long[] rnd = new long[maxIterations];
		for (int i = 0; i < maxIterations; i++) 
			rnd[i] = Math.round(Math.random()*100);
		
		DecimalFormat formatter = new DecimalFormat("0.###E0");
		String f = formatter.format(maxIterations);
		System.out.println("Beginning test on " + f + " iterations...");
		
		final long start = System.currentTimeMillis();
		for (int i = 0; i < maxIterations; i++) total += Math.sqrt(rnd[i]);
		final long stop = System.currentTimeMillis();
		
		final long elapsed = (stop - start); 
		System.out.println("Time elapsed: " + elapsed);
		System.out.println("Sum: " + total);
	}
}

Code:

using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Threading.Tasks;

namespace TimingTest
{
    class Program
    {
        static void Main(string[] args)
        {
            int maxIterations = 100000000;
            double total = 0.0;

            int[] rnd = new int[maxIterations];
            Random rng = new Random();
            for (int i = 0; i < maxIterations; i++) 
                rnd[i] = rng.Next(100);

            Console.WriteLine("Beginning test on " + maxIterations.ToString("0.###E0") + " iterations...");

            DateTime start = DateTime.Now;
            for (long i = 0; i < maxIterations; i++) total += Math.Sqrt(rnd[i]);
            double elapsed = DateTime.Now.Subtract(start).TotalMilliseconds;

            Console.WriteLine("Time elapsed: " + elapsed);
            Console.WriteLine("Sum: " + total);
            Console.Read();
        }
    }
}

Are you doing a lot of "number crunching" with your programs? If not your test doesn't reveal much about relative overhead in execution. All you're doing is essentially a few floating point calcs that have really pretty much nothing to do with how the vast majority of programs operate

Again, problem domain should dictate your language choice.

Quote

04-22-2013 , 01:14 PM

sards

old hand

Join Date: Dec 2006 Posts: 1,554

That is true. But in general the performance of C# and Java will be comparable, to the point that performance should not be the deciding factor in choosing between them. If you absolutely need the best possible performance, use C++ instead.

Quote

04-22-2013 , 03:09 PM

skier_5

Pooh-Bah

Join Date: Jan 2006 Posts: 5,443

Quote:

Originally Posted by :::grimReaper:::

This was it. I don't know much about compilers. Can anyone explain what's going on here?

You're not using the result of the square root and consequently your loop is not doing anything so the compiler completely omits it.

Be careful you're not measuring the speed of the square root function itself (unless you need to of course) as they may not be the same.

Quote

04-22-2013 , 05:34 PM

#10

:::grimReaper:::

veteran

Join Date: Jul 2010 Posts: 3,113

Quote:

Originally Posted by sards

But looking at the initial runs, seems like the Java compiler is more "smart," or may optimize execution as seen in the initial runs.

Quote

04-22-2013 , 05:37 PM

#11

:::grimReaper:::

veteran

Join Date: Jul 2010 Posts: 3,113

Quote:

Originally Posted by skier_5

Interesting. In my second test, I stored the cumulative sum of the square roots, and Java was still outperforming. It wasn't until I was taking square roots of random integers that they were equal in performance. Why?

Quote

04-22-2013 , 07:41 PM

#12

jmark

Pooh-Bah

Join Date: Dec 2003 Posts: 3,790

In C you can set the optimization flags, like -O2 and -O3 etc. I'm sure there's something similar in Java. If you want the fastest code execution for the loop:

for i = 1 to 10
total += i;

the compiler can just replace that whole for loop with:
total = 55;

Quote

04-22-2013 , 08:44 PM

#13

sards

old hand

Join Date: Dec 2006 Posts: 1,554

A lot of the time in that test is spent on unnecessary type conversion. Try changing this line:

Code:

for (long i = 0; i < maxIterations; i++)

to this:

Code:

for (double i = 0; i < maxIterations; i++)

Quote

04-22-2013 , 09:38 PM

#14

adios

Carpal \'Tunnel

Join Date: Sep 2002 Posts: 22,558

Quote:

Originally Posted by sards

A lot of the time in that test is spent on unnecessary type conversion. Try changing this line:

Code:

for (long i = 0; i < maxIterations; i++)

to this:

Code:

for (double i = 0; i < maxIterations; i++)

No don't do that. Using doubles for the loop control is not a good idea. Just use another double var. Don't use rnd () to load the array as part of the loop, load the array and then run the test. The test is bogus though, it doesn't matter much.

Quote

04-22-2013 , 10:55 PM

#15

vamozz

banned

Join Date: Oct 2012 Posts: 264

Quote:

Originally Posted by :::grimReaper:::

maybe cause the java compiler just can't optimize fair enough with random numbers

Quote

04-23-2013 , 02:54 AM

#16

:::grimReaper:::

veteran

Join Date: Jul 2010 Posts: 3,113

Quote:

Originally Posted by adios

load the array and then run the test.

That's what I did

Quote

04-23-2013 , 03:58 AM

#17

sards

old hand

Join Date: Dec 2006 Posts: 1,554

Quote:

Originally Posted by adios

No don't do that. Using doubles for the loop control is not a good idea.

I would agree if the point of this exercise were to write good code. But in fact the point is to experiment and see how various code constructs affect compiler optimizations.

Quote

04-23-2013 , 09:47 AM

#18

adios

Carpal \'Tunnel

Join Date: Sep 2002 Posts: 22,558

Quote:

Originally Posted by sards

I would agree if the point of this exercise were to write good code. But in fact the point is to experiment and see how various code constructs affect compiler optimizations.

Which brings up what is the point of this? If it is a comparison of how Java and C# handle sqrt ok. Is or a comparison of different compilers? It fails then. If it is a comparison of overhead imposed? Another fail.

Quote

04-23-2013 , 09:53 AM

#19

adios

Carpal \'Tunnel

Join Date: Sep 2002 Posts: 22,558

Quote:

Originally Posted by :::grimReaper:::

That's what I did

How much is rnd () introducing into the overall timng?

When I look at what you're doing I am a little concerned that you are drawing erroneous conclusions from the results. Hopefully you are experimenting and not making hard and fast conclusions about the languages.

Quote

04-23-2013 , 03:33 PM

#20

:::grimReaper:::

veteran

Join Date: Jul 2010 Posts: 3,113

Quote:

Originally Posted by adios

How much is rnd () introducing into the overall timng?

0. As you can see the in code (post 6), random numbers are created and stored in an array first.

Quote

04-23-2013 , 08:35 PM

#21

meekrab

veteran

Join Date: Mar 2010 Posts: 2,798

Quote:

Originally Posted by sards

A lot of the time in that test is spent on unnecessary type conversion. Try changing this line:

Code:

for (long i = 0; i < maxIterations; i++)

to this:

Code:

for (double i = 0; i < maxIterations; i++)

Actually it really isn't, changing the ints to doubles in the code in post 6 only made a ~10ms difference on my machine, which is a bit slower than OP's. Total elapsed time of about 630ms versus 640 with the 100 million casts from int to double.

Quote

04-23-2013 , 09:12 PM

#22

jjshabado

Carpal Tunnel

Join Date: Jul 2006 Posts: 22,732

Quote:

Originally Posted by adios

Which brings up what is the point of this?

Yeah, I was curious about that too. As a learning exercise its cool but if OP was hoping for useful information about the relative performance of Java vs C# it probably fails.

Quote

04-25-2013 , 06:21 PM

#23

punter11235

Carpal \'Tunnel

Join Date: Mar 2005 Posts: 8,210

Don't know about sqrt but one reason C# will be always slower than Java at number crunching is that it (current implementations) doesn't utilize some of the instructions of modern processors.
You can try googling SIMD and C# for more if you are interested in your specific case but my suspicion is that it's the reason.

Quote

04-25-2013 , 06:45 PM

#24

adios

Carpal \'Tunnel

Join Date: Sep 2002 Posts: 22,558

Quote:

Originally Posted by punter11235

I would think that if C# didn't take advantage of floating point instructions while Java did the relative time difference would be even greater.

Quote

04-25-2013 , 07:11 PM

#25

punter11235

Carpal \'Tunnel

Join Date: Mar 2005 Posts: 8,210

I am clueless when it comes to details as I don't really do C# but I remember even very simple number crunching loops written in C gained a lot after enabling SSE2/4 comparing to C# equivalent. I remember doing some research (googling) back then and gaining the impression that the situation is better in Java.
You might be right the the difference should be bigger in case of sqrt. I don't know, just wanted to point possible reason for that whch I encountered in my experience.

This btw sounds like great material for StackOverflow question. Some smart people there or sure know the answer and will give it to you for mere upvote