BigInteger on 64 bit platform

[rmcnutt]

Does anyone know if java.math.BigInteger performs significantly better
on an AMD Athlon 64 processor. I am using an AMD Athlon XP 2K. While
I'm happy with the performance, I have to wonder if the 64 is worth
the upgrade.

 

[Neil Campbell]

Does anyone know if java.math.BigInteger performs significantly better
on an AMD Athlon 64 processor. I am using an AMD Athlon XP 2K. While
I'm happy with the performance, I have to wonder if the 64 is worth
the upgrade.

Probably not - Java hasn't been properly ported to the AMD64 architecture as
yet. Blackdown have released a version for Athlon64/Opteron CPUs, but I
believe the JVM is still essentially a 32 bit application, so it will
perform as well as on any other processor.

Having said that, the Athlon64 processors are pretty quick anyway, so you
will see an increase; but probably not much more than you would get from
any higher end 32-bit CPU.

 

[Juergen Kreileder]

Probably not - Java hasn't been properly ported to the AMD64
architecture as yet. Blackdown have released a version for
Athlon64/Opteron CPUs, but I believe the JVM is still essentially a
32 bit application, so it will perform as well as on any other
processor.

No it's a real 64-bit AMD64 application which takes advantage of the
increased register number and size, etc. (Sun's 1.5 VM for AMD64 is
based on a our code BTW.)

Note that there's no reason to make a special 32-bit VM for AMD64: The
32-bit mode is compatible with x86, it runs x86 VMs just fine. As for
performance: The 64-bit AMD64 VM is faster than the 32-bit x86 VM
running on AMD64 in allmost all benchmarks (mostly due to the
increased register number; code using Java longs also benefits a lot
from having 64-bit wide registers).


Juergen

 

[Neil Campbell]

No it's a real 64-bit AMD64 application which takes advantage of the
increased register number and size, etc. (Sun's 1.5 VM for AMD64 is
based on a our code BTW.)

Note that there's no reason to make a special 32-bit VM for AMD64: The
32-bit mode is compatible with x86, it runs x86 VMs just fine. As for
performance: The 64-bit AMD64 VM is faster than the 32-bit x86 VM
running on AMD64 in allmost all benchmarks (mostly due to the
increased register number; code using Java longs also benefits a lot
from having 64-bit wide registers).

Sorry, I didn't realise this was a full blown 64-bit app. I thought that as
Sun's wasn't out until the 1.5 release, Blackdown would follow suit.
Thanks for clearing this up.

I use the AMD64 version at work on an Opteron, to my mind it doesn't seem
significantly faster than the 32bit equivalents. The main reason I use
this VM is that I had experienced some problems with earlier versions on
that system, in particular with apps using AWT (this may have been due to
some peculiarity of my X installation, but the problems went away when I
upgraded so I didn't investigate further).

 

[Juergen Kreileder]

I use the AMD64 version at work on an Opteron, to my mind it doesn't
seem significantly faster than the 32bit equivalents.

Note that the AMD64 version always uses the HotSpot Server VM, there's
no HotSpot Client VM for AMD64.
The IA32 version uses the HotSpot Client VM by default.
So if you do performance comparisons, make sure you compare the AMD64
VM with the IA32 Server VM ("java -server ...").

The Client VM performs a bit better than the Server VM for GUI
applications (at least initially) although the Server VM is
technically superior.


Juergen

 

[Neil Campbell]

Note that the AMD64 version always uses the HotSpot Server VM, there's
no HotSpot Client VM for AMD64.
The IA32 version uses the HotSpot Client VM by default.
So if you do performance comparisons, make sure you compare the AMD64
VM with the IA32 Server VM ("java -server ...").

OK, fair enough. I haven't done any benchmarking as such, but the reason
you identified is probably why I'm not seeing the sorts of gains one might
expect; I am indeed using primarily client apps (although they do tend to
run for several days at a time).

To be honest though, my Opteron box is so much faster than my previous
computer even in 32-bit mode that it would be difficult to make many
meaningful comparisons in my case.

Is a 'Client Mode' VM for AMD64 in development?

Cheers,

 

[rmcnutt]

OK, fair enough. I haven't done any benchmarking as such, but the reason
you identified is probably why I'm not seeing the sorts of gains one might
expect; I am indeed using primarily client apps (although they do tend to
run for several days at a time).

To be honest though, my Opteron box is so much faster than my previous
computer even in 32-bit mode that it would be difficult to make many
meaningful comparisons in my case.

Is a 'Client Mode' VM for AMD64 in development?

Cheers,

Thanks for the the information. I was not familiar with Blackdown.
I've been using Sun's SDK 1.4. I'm using Fedora Core at present.
Will Blackdown work Fedora64?

Regards,

 

[Juergen Kreileder]

I'm sure you'll understand that I only can comment this from the
Blackdown standpoint: We haven't decided yet if we'll develop a
HotSpot Client VM for AMD64.

Thanks for the the information. I was not familiar with Blackdown.
I've been using Sun's SDK 1.4. I'm using Fedora Core at present.
Will Blackdown work Fedora64?

Yes.

As for your original question (so you don't have to post it daily now ;-),
write a benchmark testing operations that are important for you
and I'll give it try. (I've done a simple benchmark with
BigInteger::multiply(), the AMD64 VM was about 25% faster than the
IA32 VM on the same machine. But this result may have very limited
meaning for your actual code.)


Juergen

 

[Neil Campbell]

I'm sure you'll understand that I only can comment this from the
Blackdown standpoint: We haven't decided yet if we'll develop a
HotSpot Client VM for AMD64.

OK. Thanks a lot for all the info!