benchmarks? java vs .net

[King Raz]

The shootout site has benchmarks comparing different languages. It
includes C# Mono vs Java but not C# .NET vs Java. So I went through
all the benchmark on the site ...

http://kingrazi.blogspot.com/2008/05...enchmarks.html

Just to keep the post on topic for my friends at comp.lang.c++, how do
I play default windows sounds with C++?

[Razii]

On Tue, 03 Jun 2008 20:07:00 +0100, Jon Harrop <jo*@ffconsulta ncy.com>
wrote:

>So having to write your own basic trig functions in Java so that you can be
only 2x slower than other languages doesn't bother you?

It's not 2x slower with my trig function on my computer.

On the other hand, C# results are not accurate for trig functions.

using System;
using System.IO;

class Test
{
static void Main(){
Console.WriteLi ne(Math.Sin (1e15));
}
}
The answer I get from C# is: 0.8582721324763 73

with Java I get 0.8582727931702 359

Checking the answer with maple, I get
0.8582727931702 358355238863908 484066466002034

How are you going to fix C# in this case?

Does it bother you that C# gave wrong answer?

[Razii]

On Tue, 3 Jun 2008 20:11:45 +0100, Jon Skeet [C# MVP]
<sk***@pobox.co mwrote:

>Now, you're using *those results* to form conclusions, right? If not,
there was little point in posting them. However, those results are of
Mono, not .NET.

Is this guy Jon Skeet really this stupid?

Let me know so I can add him to ignore list.

[Mark Thornton]

Jon Skeet [C# MVP] wrote:

Mark Thornton <mt*******@optr ak.co.ukwrote:

<snip>

>There are some FJ examples here:
http://artisans-serverintellect-com....efault.asp?W32

The recursive matrix multiplication is my contribution.

Right. It still looks a bit involved - which is only to be expected,
really. I haven't taken it in fully yet though - will aim to do so at a
later date. Might try porting to Parallel Extensions...

Frameworks like Fork/Join can only go so far in easing tasks like matrix
multiplication. Given that my Java code won't be making best use of
SSE, I was quite pleased with the performance of the result.

Mark Thornton

[Barry Kelly]

Razii wrote:

Is this guy Jon Skeet really this stupid?

Let me know so I can add him to ignore list.

Razii, if you hope to ever have any positive effect in your
communication - e.g. convince anyone that you have a clue - you'll have
to work much harder not to look like a naive 13 year old with
undeveloped social skills.

Plonk etc.

-- Barry

--
http://barrkel.blogspot.com/

[Barry Kelly]

Jon Skeet wrote:

Now using Parallel Extensions, and the lamdba expressions of C# 3:

public override void Generate()
{
Parallel.For(0, Height, row =>
{
int index = row * Width;
for (int col = 0; col < Width; col++)
{
Data[index++] = ComputeMandelbr otIndex(row, col);
}
});
}

I can't imagine many transformations being simpler than that - and the
results are great.

How have you found Parallel.For to perform?

I've found that my own naive (i.e. quick hack with little thought)
implementations of Parallel.For that dispatch over a fixed number of
threads (one thread per core) outperform the ones in the TPL by quite a
bit - TPL seemed to have had around ~20% overhead in the particular
microbenchmark I was trying to scale over my quad core.

I must dig it out and analyse the discrepancy, and blog about it, when I
find time.

-- Barry

--
http://barrkel.blogspot.com/

[Razii]

On Tue, 03 Jun 2008 22:27:12 +0100, Barry Kelly
<ba***********@ gmail.comwrote:

>Plonk etc.

You must be a newbie, moron. No one ever plonks me. You can try
though, as you did.

As for Jon Skeet, what kind of reply do you expect? I explained to him
that I don't have Mono, and the benchmarks I am using are at shootout
site (not that I myself am using Mono). The link to the blog I posted
also makes that clear, but yet the guy continued with the same line
that I am comparing Mono. Is he stupid or just doesn't take the time
to read carefully? Take your pick.

[Jon Skeet [C# MVP]]

Barry Kelly <ba***********@ gmail.comwrote:

I can't imagine many transformations being simpler than that - and the
results are great.

How have you found Parallel.For to perform?

Very well - see my last few blog posts.

I've found that my own naive (i.e. quick hack with little thought)
implementations of Parallel.For that dispatch over a fixed number of
threads (one thread per core) outperform the ones in the TPL by quite a
bit - TPL seemed to have had around ~20% overhead in the particular
microbenchmark I was trying to scale over my quad core.

I must dig it out and analyse the discrepancy, and blog about it, when I
find time.

I wouldn't be surprised if some individual test cases did badly in
microbenchmarks . Without even looking at your code, my guess is that
you may find your code doesn't work as well when there's other load on
the processor (e.g. effectively taking up one core with a different
process) whereas work stealing should (I believe) cope with that
reasonably well.

But as I point out on the most recent blog entry, accurately
benchmarking this kind of thing and being able to interpret the results
sensibly is basically beyond my current understanding (and time to
spend on it).

--
Jon Skeet - <sk***@pobox.co m>
Web site: http://www.pobox.com/~skeet
Blog: http://www.msmvps.com/jon.skeet
C# in Depth: http://csharpindepth.com

[Razii]

On Tue, 3 Jun 2008 21:04:47 +0100, Jon Skeet [C# MVP]
<sk***@pobox.co mwrote:

>Has it occurred to you that that may have an effect on the results?

No, it won't have a major effect.

[Razii]

On Tue, 03 Jun 2008 15:14:52 -0500, Razii <kl*****@mail.c omwrote:

>The answer I get from C# is: 0.8582721324763 73

with Java I get 0.8582727931702 359

Checking the answer with maple, I get
0.858272793170 235835523886390 848406646600203 4

How are you going to fix C# in this case?

Does it bother you that C# gave wrong answer?

Console.WriteLi ne(Math.Sin (1e7));

0.4205477931907 71 (C# with .NET)
0.4205477931907 824912985065897 4095 (right answer)

Console.WriteLi ne(Math.Sin (1e10));

-0.4875060250762 7 (C# with .NET)
-0.4875060250875 106915277942943 4811 (right answer)

Console.WriteLi ne(Math.Sin (1e15));

0.8582721324763 73 (C# with .NET)
0.8582727931702 358355238863908 484 (right answer)

So C# doesn't get 15-17 digit accuracy of double for sin and cos.
That's the only reason it's faster in partialsums benchmark. By using
FastMath class, the times for both is about same on my computer.

On the other hand, you still need to 'optimize' 4 benchmarks where
..NET is much slower: binarytrees, sum-col, recursive, revcomp.

[Lew]

Mark Thornton wrote:

Given that my Java code won't be making best use of SSE

Are you sure of that?

If true, it's one more area for Hotspot authors to approach.

--
Lew