[dangermouse_197]

Hello I must admit that I am new to the Apple world but did in fact buy an iBook G4 800MHz late last year and took the RAM up to 640MB. I sold it about 2 months later (maybe a little less) as I found it just too slow...I was going to use it for development work but found compiling just too slow.

And here I am today missing my Apple and am back to Win32 machines (sorry...no flames please). What I want to know is, if I sprung for either a 2nd hand 867MHz PB 12" or a new 1GHz PB 12" would I see an increase in performance ?? If I would, is there much difference between the 867MHz and 1GHz in terms of 'noticeable' speed (i.e. were other components changed that increased performance).

I'm not all that bothered about 3D rendering performance as I use my 9800Pro equipped desktop for that but I really liked the MacOS...just the speed of the 800MHz G4 left me slightly unimpressed.

If anyone can offer any advice, I would appreciate your thoughts.

[Lateralus]

If your budget is low, and you don't need portability, I think you should go for a refurbished Dual 1.25GHz Power Mac G4. I've seen them for around $1,499, and seen refurbished Single 1.25GHz models for $1,099.

I don't think you'd find an 867MHz 12" PowerBook much faster than your iBook.

[coolmacdude]

Originally posted by dangermouse_197:
I really liked the MacOS...just the speed of the 800MHz G4 left me slightly unimpressed.

But see that's exactly it. Once you get used to OS X, any speed disadvantage your computer has compared to a PC will be repayed many times over by the huge boost in productivity and ease of use you will get.

It really is true that the computer user is the greatest bottleneck in any workflow, not any hardware component. Doing ordinary things is so much more complicated than it needs to be on Windows. If you give the Mac a serious chance, you won't be disappointed.

[cambro]

Originally posted by dangermouse_197:
I want to know is, if I sprung for either a 2nd hand 867MHz PB 12" or a new 1GHz PB 12" would I see an increase in performance ??

The Powerbook has twice the L2 cache (512 K) that the iBook has. This definitely translates into better performance on G4 systems.

This only applies to the Ghz 12" PB, though. Don't get the 867.

[Dr.Michael]

Hi dangermouse_197,

I must repeat what I have written several times (and never got a reply):
As a developer I use a powerbook 12 inch/1GHz. It IS slower than my IBM centrino 1.6 GHz, but its much more comfortable to work with.

The missing speed is an ADVANTAGE for developers. You simply see if your code is slow. It would be a very bad thing if your clients, who may use old computers, are the first who find slow code.

And if an iBook G4 compiles too slow for you, you either have really big projects or something is wrong with your code.

So buy a Powerbook and enjoy!

Michael

[Pierre B.]

Originally posted by Dr.Michael:

So buy a Powerbook and enjoy!

And to not forget: not the Powerbook 12" 867 MHz, since its processor, apart the 67 MHz difference in clock frequency, is virtually the same as the one in the iBook you had (256 KB L2 cache).

[Pierre B.]

Originally posted by Dr.Michael:
As a developer I use a powerbook 12 inch/1GHz. It IS slower than my IBM centrino 1.6 GHz, but its much more comfortable to work with.

Just curious, have you tried to run the same code on the two machines to compare them? For example calculate recursively big factorials or run some other processor intensive algorithms? It would be interesting to see here results like that.

[SEkker]

Most iBooks are limited by the speed of their HD. A stock 4200 rpm HD will be a major bottleneck for all work in OSX. I am still amazed at how fast my wife's G3 400 MHz Pismo is now that I dropped in a 45 GB 5400 HD.

If you want to get back in the mac world and do not need a portable, get a cheaper G5 desktop -- it will be more than fast enough.

If you really need speed in a portable, puchase a BTO PB G4, 1 GHz minimum. There are 7200 rpm HDs that may not be available from Apple, if you are comfortable with after puchase modifications to your Mac. Otherwise, be sure to get a 5400 rpm HD from Apple. And 1 GHz RAM.

I bet if you had put a 7200 rpm HD in your old iBook, you would not have felt compelled to sell it.

[Dr.Michael]

Originally posted by Pierre B.:
Just curious, have you tried to run the same code on the two machines to compare them? For example calculate recursively big factorials or run some other processor intensive algorithms? It would be interesting to see here results like that.

Hi Pierre,

I have never made clean investigations. I simply run my (Java) code on both machines. I can also compare OS X and YellowDog Linux.

My test program is mainly a big decimal iterative pi calculation. But it does not only test the raw processor speed. As I saw from the upgrade from .2 to .3 the implementation of the apple virtual machine has significant effects.

10.2 -10.3: 50 % faster
Thinkpad under suse 9.0: 30 % faster than 10.3
YellowDog with a ibm ppc virtual machine: 30 % slower than 10.3

[Person Man]

Originally posted by SEkker:
Most iBooks are limited by the speed of their HD. A stock 4200 rpm HD will be a major bottleneck for all work in OSX. I am still amazed at how fast my wife's G3 400 MHz Pismo is now that I dropped in a 45 GB 5400 HD.

If you want to get back in the mac world and do not need a portable, get a cheaper G5 desktop -- it will be more than fast enough.

If you really need speed in a portable, puchase a BTO PB G4, 1 GHz minimum. There are 7200 rpm HDs that may not be available from Apple, if you are comfortable with after puchase modifications to your Mac. Otherwise, be sure to get a 5400 rpm HD from Apple. And 1 GHz RAM.

I bet if you had put a 7200 rpm HD in your old iBook, you would not have felt compelled to sell it.

On the new Aluminum PowerBooks, hard drives are not a customer-installable part like they were on earlier models. Open up the machine yourself to put the new drive in and you void the warranty. You'd have to take it to an Apple authorized service provider.

[Pierre B.]

Originally posted by Dr.Michael:

My test program is mainly a big decimal iterative pi calculation. But it does not only test the raw processor speed. As I saw from the upgrade from .2 to .3 the implementation of the apple virtual machine has significant effects.

10.2 -10.3: 50 % faster
Thinkpad under suse 9.0: 30 % faster than 10.3
YellowDog with a ibm ppc virtual machine: 30 % slower than 10.3

Thanks for taking the time for this test! Yes, having a virtual machine in the middle does a different test, however I am a bit surprised that the 60% advantage in clock speed of the Centrino is translated to 30% advantage in execution speed. In this test the 1 GHz Powerbook behaves like a 1.25 GHz Centrino.

Is it possible to you to write and run the same code in C and see what of a difference that makes? I believe that the Centrino would show a much more pronounced advantage, as the G4 is not very strong in FP calculations. But then again there is the compiler issue...

[Dr.Michael]

Is it possible to you to write and run the same code in C and see what of a difference that makes? [/B]

No, not at the moment.
Its years ago that I have touched c.
But feel free... I would be interested to see the results.

[Pierre B.]

Originally posted by Dr.Michael:
No, not at the moment.
Its years ago that I have touched c.

You can also try to compile a pi calculation program based on FFT and AGM from here. When asked, you could enter a FFT length of 200000 to calculate 1048576 digits of pi.

But feel free... I would be interested to see the results.

Although I don't know Java, I would be interested to see your code. Is it a problem to post it here or to post a link to download it?

By the way, piX is a nice and rather efficient program with GUI to calculate arbitrary number (more than 100) of pi digits.

[Dr.Michael]

Originally posted by Pierre B.:

Although I don't know Java, I would be interested to see your code. Is it a problem to post it here or to post a link to download it?

I had to make up my mind and consult my organizer before I could answer.

Yes, it would be possible to post the code. But I have to extract the core calculation and translate my comments into english.

I see if I find time this evening. Since this is a performance thread - why not.

You all please keep in mind that big decimal operations are only one thing you can do. Its by no means representative for overall performance. But its one of the worst things you can do with java. So if this performs well, its a sign, that apples java team has done a good job (they have :0).

Maybe there is someone out there who can translate my code to c or delphi or whatever. I would appreciate this.

So stay tuned. I'll be back soon.

Michael

[Dr.Michael]

Originally posted by Pierre B.:

Although I don't know Java, I would be interested to see your code. Is it a problem to post it here or to post a link to download it?

Ok, here is the code.

How can you compile it?

create a folder
create a text file. The name MUST be "Pi.java"
copy everything between //begin of code and //end of code into the file and save it.

open the terminal
cd to the folder (or type "cd " and drag the folder into the terminal)

check if java is installed by typing
"java -version" (then hit return)

to compile type "javac Pi.java"

this creates a file called "Pi.class"
run it with "java -cp ./ Pi"

My score:
On my Powerbook (12 inch, 1GHz, 768 RAM)
OS 10.3: 310 ms
Yellow Dog Linux 3.01: 485 ms
Windows NT4 in VirtualPC 6.1: 540 ms

I'll add the IBM Thinkpad times tomorrow.

Happy compiling,

Michael


//begin of code
/**
* Pi.java
* Description: simple recursive Pi calculation using big decimals
* @author Michael Wetzstein
*/


import java.math.BigDecimal;

public class Pi {

//scale is the measure of numbers
//behind the decimal point
private static int scale = 100;

//loopsize is the number of loops after
//which the timestamp is taken
private static String loopsize = "10000";

//variables for time measurement
private long start;
private long time;

//variables for pi calculation
private BigDecimal pi = new BigDecimal(0.0).setScale(scale, BigDecimal.ROUND_HALF_EVEN);
private static BigDecimal one = new BigDecimal("1.0").setScale(1, BigDecimal.ROUND_UNNECESSARY);
private static BigDecimal minusone = new BigDecimal("-1.0").setScale(1, BigDecimal.ROUND_UNNECESSARY);
private static BigDecimal two = new BigDecimal("2.0").setScale(1, BigDecimal.ROUND_UNNECESSARY);
private BigDecimal numerator = new BigDecimal("-4.0").setScale(1, BigDecimal.ROUND_HALF_EVEN);
private BigDecimal denominator = new BigDecimal("0.0").setScale(1, BigDecimal.ROUND_HALF_EVEN);
private static BigDecimal counter = new BigDecimal(loopsize).setScale(1, BigDecimal.ROUND_UNNECESSARY);
private BigDecimal a = new BigDecimal("0.0").setScale(1, BigDecimal.ROUND_UNNECESSARY);
private BigDecimal i = new BigDecimal("0.0").setScale(1, BigDecimal.ROUND_UNNECESSARY);


public static void main (String arguments[]) {
new Pi();
}

public Pi(){

//outer loop: only for time measurement
do {
a=a.add(counter);
start = System.currentTimeMillis();

//inner loop crunches the numbers
do {
i=i.add(one);
denominator=(two.multiply(i)).subtract(one);
denominator = denominator.setScale(1, BigDecimal.ROUND_UNNECESSARY);
numerator=numerator.multiply(minusone);
numerator=numerator.setScale(1, BigDecimal.ROUND_UNNECESSARY);
pi= pi.add(numerator.divide(denominator, scale, BigDecimal.ROUND_HALF_EVEN));
}
while (i.compareTo(a) < 0);

time = System.currentTimeMillis() - start;

System.out.print("Iteration no: " +i.longValue() +". Time for " +loopsize
+" iterations: " +time +" ms. Pi: " +pi.setScale(50, BigDecimal.ROUND_HALF_EVEN).toString() );
System.out.print((char)13);
}
while (true); //forever.
}
}
//end of code

[Dr.Michael]

Originally posted by Dr.Michael:


My score:
On my Powerbook (12 inch, 1GHz, 768 RAM)
OS 10.3: 310 ms
Yellow Dog Linux 3.01: 485 ms
Windows NT4 in VirtualPC 6.1: 540 ms

I'll add the IBM Thinkpad times tomorrow.

OK, here are the PC scores:
IBM Thinkpad T40 (1.6 GHz Pentium M, 1 GB RAM)
Suse Linux 9.0: 112 ms
Windows XP: 105 ms

So again, if someone can provide platform independent c code I would appreciate that.
And dangermouse, now you have a confirmation that your impression wasn't too false :o).

So please everybody post your times.

BTW: the times are milliseconds that are needed to do 10.000 loops of Pi calculation.

Michael

[Pierre B.]

Originally posted by Dr.Michael:
Ok, here is the code.

Great, thanks! I observe that your calculation makes use of special Java libraries for arbitrary decimal precision. I am not sure if it is simple to port that in C.

Anyway, I would suggest to try the program of the other link I posted before. Just download the pi_fftc6_src.tgz (52KB) package and read the instructions.They provide the source code for Unix systems (there is a Makefile which makes compilation trivial), as well as the binaries for Windows. So, it is very easy to install and run it. You have only to provide a FFT length (200000 is for a little more than 1000000 digits), and at the end the program prints out the execution time as well as the calculated pi! At this moment it does not make sense for me to post numbers as my Powerbook is under heavy load (100% CPU), and it will remain so for some hours.

A "strings" on the windows executable reveals that it has been compiled with Intel(R) C++ Compiler for 32-bit applications, Version 5.0.1 Build 010525Z.

It would be very interesting to see relative performance numbers alongside the ones you provided with your own code. I think the compilers too would make a great difference.

EDIT: sorry, the correct link with the source code and the Windows binaries is this one.

[Pierre B.]

Originally posted by Dr.Michael:
OK, here are the PC scores:
IBM Thinkpad T40 (1.6 GHz Pentium M, 1 GB RAM)
Suse Linux 9.0: 112 ms
Windows XP: 105 ms

I don't understand something here. You said before:

Thinkpad under suse 9.0: 30 % faster than 10.3

In the new test the Thinkpad appears about 200% faster than the Powerbook under Panther .

[Dr.Michael]

Originally posted by Pierre B.:
I don't understand something here. You said before:

In the new test the Thinkpad appears about 200% faster than the Powerbook under Panther .

Correct.
I have modified my program and optimized it a little. This had more impact on the pc speed.

I have rewritten my little program again to make it more c-compatible. Instead of BigDecimals I have used double and long.
This makes only limited sense in terms of Pi calculation but is easier to copy and avoids implementation details in the libraries.

The times for 1.000.000 iterations now:
Thinkpad/Linux 48 ms
Thinkpad/XP 50 ms

I give the Powerbook performance this evening. And I will try your link then.

Here is the modified code. Please copy it into a file called Pi_double.java

//begin code
/**
* Pi_double.java
* Description: simple recursive Pi calculation using big decimals
* @author Michael Wetzstein
*/


public class Pi_double {

//loopsize is the number of loops after
//which the timestamp is taken
private static long loopsize = 1000000L;

//variables for time measurement
private long start;
private long time;

//variables for pi calculation
private double pi = 0.0d;
private double numerator = -4.0d;
private long a = 0L;
private long i = 0L;

public static void main (String arguments[]) {
new Pi_double();
}

public Pi_double(){

//outer loop: only for time measurement
do {
a+=loopsize;
start = System.currentTimeMillis();

//inner loop crunches the numbers
do {
i+=1L;
numerator*= -1.0d;
pi+= numerator/(2*i -1);
}
while (i<a);

time = System.currentTimeMillis() - start;

System.out.print("Iteration no: " +i +". Time for " +loopsize
+" iterations: " +time +" ms. Pi: " +pi );
System.out.print((char)13);
}
while (true); //forever.
}
}
//end code

[Pierre B.]

Originally posted by Dr.Michael:
I have rewritten my little program again to make it more c-compatible. Instead of BigDecimals I have used double and long.
This makes only limited sense in terms of Pi calculation but is easier to copy and avoids implementation details in the libraries.

OK, here is a C version of the above. You can save the file as, say, Pi_double.c, and for better performance, you can compile the file with the -O3 option:

cc -O3 Pi_double.c <enter>

You start execution typing

./a.out <enter>

The program runs nonstop, you have to type Ctrl-C to interrupt it in the terminal.

/* BEGIN OF CODE */

#include <stdio.h>
#include <sys/types.h>
#include <sys/times.h>

#define CLK_TCK 100

main(){
int a, i, j;
struct tms before, after;
float numerator;
float pi;

numerator = -4.0;
a = 1000000;
j = 0;

while( j < a ){

pi = 0.0;
times( &before );

for( i = 1; i <= a; i++ ){

numerator = numerator * ( -1.0 );
pi = pi + numerator / ( 2 * i - 1 );

}

times( &after );

printf( "User time: %ld milliseconds\n", 1000 * ( after.tms_utime - before.tms_utime ) / CLK_TCK );
printf( "Pi = %.20f\n\n", pi );

}

}

/* END OF CODE */

Make sure that the printf instructions are arranged in one line each. If do you want more iterations, you can set the variable "a" at a higher value (the program will run for a = 1000000)

[Dr.Michael]

Great!
This will make a good comparison.
I'll try it this evening.

What are your results?

[Pierre B.]

Originally posted by Dr.Michael:
What are your results?

12" Powerbook, 867 MHz, 10.2.8 with all system updates applied.

Your new java code: ~222 ms.
C code (with -O3 option): 50-70 ms.

Always a = 1000000.

[fibroptikl]

Originally posted by Dr.Michael:
OK, here are the PC scores:
IBM Thinkpad T40 (1.6 GHz Pentium M, 1 GB RAM)
Suse Linux 9.0: 112 ms
Windows XP: 105 ms

So again, if someone can provide platform independent c code I would appreciate that.
And dangermouse, now you have a confirmation that your impression wasn't too false :o).

So please everybody post your times.

BTW: the times are milliseconds that are needed to do 10.000 loops of Pi calculation.

Michael

I tried to e-mail you, but you don't allow for e-mails and PM's are not allowed either to see if you had any pictures of your IBM ThinkPad T40; I was doing some research on them.

[Dr.Michael]

Originally posted by Pierre B.:
12" Powerbook, 867 MHz, 10.2.8 with all system updates applied.

Your new java code: ~222 ms.
C code (with -O3 option): 50-70 ms.

Always a = 1000000.

Here are my results:

c-code:
Thinkpad (Linux): 40 ms
Powerbook: 50 ms (40-60)

new java-code:
Thinkpad (Linux): 48 ms
Powerbook: 202 ms

I could tune a little more for the Powerbook (for loop is faster than do while), so I got it down to 175 ms.

Ok, the c-scores save the day!

But I definitely have to take back what I said about Apples Java implementation.

We see that java is only 20% slower on the Thinkpad. This is what I have read in several tests. And it is valid for c-code compared to java. In object oriented software projects java is as good or better than c++.

But the difference on the mac is disturbing. c is 4 times as fast compared to java. And, I must say, java guis are damn slow on the mac. Especially if they are used in aqua mode and not in native java metal look.

So, I am sorry, dangermouse was completely right and I was wrong.

And Apple please put a little more manpower into your java team.

Thanks for this great test, Pierre.
And all the others: please do some tests on your machines (I am curious to see iBook and 1.33 GHz Powerbook scores). Its very easy to compile. Please try.

[Dr.Michael]

Originally posted by fibroptikl:
I tried to e-mail you, but you don't allow for e-mails and PM's are not allowed either to see if you had any pictures of your IBM ThinkPad T40; I was doing some research on them.

I am sorry, I have no pictures of the Thinkpad and I have no digital camera to take some.

But the Thinkpad looks like any other Thinkpad, nothing special. Look for pictures of the T41 on the web. The two are very much alike.

Michael

[Pierre B.]

I just tested the pi_fftc6 code in my Powerbook, 12" 867 MHz, 10.2.8 and in particular the pi_cs binary. Here are the results, for FFT length 200000 (it calculates 1048576 digits of Pi):

Under 10.2.8: 35 sec.
Under VPC 6.1, Win2k (~530 MHz Pentium): 165 sec.

I repeat that installing this program is pretty trivial: after decompressing the zip file, you go from the terminal in the directory created and you simply type

make <enter>

If you want to delete the object files, you can say

make clean <enter>

The program "dgt_div" created after the installation (a mere compilation actually), can be used to arrange the digits in the result file (pi.dat) in a more usable format. Try it and you will see. You will finally obtain more than 1000000 Pi digits in a text file and in a very readable format!

[Pierre B.]

Originally posted by Pierre B.:
I just tested the pi_fftc6 code in my Powerbook, 12" 867 MHz, 10.2.8 and in particular the pi_cs binary. Here are the results, for FFT length 200000 (it calculates 1048576 digits of Pi):

Under 10.2.8: 35 sec.
Under VPC 6.1, Win2k (~530 MHz Pentium): 165 sec.

New test of the pi_fftc6 code:

Dell Latitude D800, Intel Pentium M Processor, 1.60 GHz, WinXP: 13 sec.

If one takes into account the pale G4 bus (133 MHz for the 867 MHz Powerbook vs. 400 MHz for the Pentium M) and its anemic L2 cache (256 KB for the Powerbook vs. 1 MB on the Pentium M), I would say that the 867 MHz G4 performs surpisingly well in this test for its clock speed. And this ignores the Altivec unit.

Any 15" or 17" Powerbook users to try this test?

[Pierre B.]

Originally posted by Pierre B.:
New test of the pi_fftc6 code:

One more, probably the last from me for this code:

Powerbook G3 266 MHz (Wallstreet), MacOS X 10.1.5: 92 sec.

Not bad (I mean compared to the 530 MHz Pentium posted previously), given that this machine has only external L2 cache (1 MB) and a 66 MHz bus. Oh, yeah, its processor is only 266 MHz.