math.BigInteger Performance

[rmcnutt]

Does anyone know if the AMD Athlon 64 offers a significant performance
increase for the BigInteger class? I'm currently using an Athlon XP
2K which is faster than a comparable Intel for Integer and Floating
Point. I'm wondering if the the 64 is worth the money.

 

[John C. Bollinger]

Does anyone know if the AMD Athlon 64 offers a significant performance
increase for the BigInteger class? I'm currently using an Athlon XP
2K which is faster than a comparable Intel for Integer and Floating
Point. I'm wondering if the the 64 is worth the money.

I'm sorry, I thought you wrote "BigInteger" and "performance" in the
same sentence. Oh, wait, you did! How bizarre.

I don't actually know whether you would see an improvement with an
Athlon 64, but my bet would be no -- until and unless you get a 64-bit
JVM and a 64-bit OS to run it on. Perhaps you might get an overall
boost from the general 32-bit mode performance of the '64, but for that
you might also consider an Athlon XP 3K, or whatever they're up to now.
You might even be able to use that in your current computer.


John Bollinger
(e-mail address removed)

 

[Mark Thornton]

I'm sorry, I thought you wrote "BigInteger" and "performance" in the
same sentence. Oh, wait, you did! How bizarre.

I don't actually know whether you would see an improvement with an
Athlon 64, but my bet would be no -- until and unless you get a 64-bit
JVM and a 64-bit OS to run it on. Perhaps you might get an overall
boost from the general 32-bit mode performance of the '64, but for that
you might also consider an Athlon XP 3K, or whatever they're up to now.
You might even be able to use that in your current computer.

Are you aware that there IS a 64 bit JVM and 64 bit OS available NOW for
the AMD64? (Blackdown and Linux respectively.)

Mark Thornton

 

[John C. Bollinger]

Are you aware that there IS a 64 bit JVM and 64 bit OS available NOW for
the AMD64? (Blackdown and Linux respectively.)

I was aware that 64-bit Linux was available for AMD64. I was not aware
that Blackdown had a 64-bit optimized version of their JVM for that
platform. That's pretty cool.

It remains an open question whether that combination would be a
satisfactory alternative for the OP, but in any event I am curious as to
just what kind of performance differences one would see with it. Does
Blackdown make any claims about it? Can anyone out there comment?


John Bollinger
(e-mail address removed)

 

[Juergen Kreileder]

I was aware that 64-bit Linux was available for AMD64. I was not
aware that Blackdown had a 64-bit optimized version of their JVM for
that platform. That's pretty cool.

It remains an open question whether that combination would be a
satisfactory alternative for the OP,

It depends on what his code does. I've done a simple test with
BigInteger::multiply(), the 64-bit VM was about 25% faster. But I
wouldn't claim all BigInteger code will run 25% faster based on that
single test.

but in any event I am curious as to just what kind of performance
differences one would see with it. Does Blackdown make any claims
about it? Can anyone out there comment?

It's noticeably faster than the IA32 VM running on the same machine in
_almost_ all benchmarks. The main reason is that you have more
registers in 64-bit mode (16 vs 8 GPRs, 16 vs 8 SSE registers).
Code using Java longs also benefits a lot (ie. more than other code)
from having 64-bit wide registers.
(These results are pretty much consistent with those of GCC benchmarks
comparing 32-bit and 64-bit code on AMD64.)

Note that the AMD64 VM is HotSpot-Server only. So if you do
benchmarks against the 32-bit VM, make sure you compare it to the IA32
Server VM (java -server ...).


Juergen