RE: [Jython-users] slow JPython math on HP??

[Tim P. <ti...@ho...>]

[attribution lost]
> HP:
>
> JPython 1.1+09 on java1.2.2.05 (JIT: null)
> 86.5       <--  interpretted, very slow
>
> JPython 1.1+09 on java1.2.2.05 (JIT: HP)
> 11.8     <-- classic JIT
>
> JPython 1.1+09 on java1.2.2.05 (JIT: null)
> 14.9     <--  HotSpot JIT

[Finn Bock]
> My numbers are on Win2k & JDK1.3 & 300Mhz P3:
>
> Jython 2.0beta2 on java1.3.0 (JIT: null)
> 37.74500000476837  <-- interpretted
>
> Jython 2.0beta2 on java1.3.0 (JIT: null)
> 4.736999988555908  <-- hotspot
>
> While somewhat faster then your HP it is in the within the same range.

[AL]
> Linux        512 MHz   dual Celeron (brain dead?) Pentium
>
> Linux: /home/mudd[1]  jpython
> JPython 1.1 on java1.2.2-RC2 (JIT: javacomp)
> Copyright (C) 1997-1999 Corporation for National Research Initiatives
> >>> from time import time
> >>> t1=time(); x=2L**10000; delay=time()-t1; print delay
> 0.00499        <-- About 2,400 times faster than HP's best numbers??!?
> >>>

[FB]
> That's fast! Does it actually calculate the correct value?
>
> The power calculation is done fully by the java.math.BigInteger
> implementation. The BigInteger.pow() on JDK1.3 (and HP too I
> expect) is implemented in java as some kind of loop which I
> don't understand.
>
> If the linux JVM are doing its BigInteger stuff in native code, that
> might explain the difference.

Let's try it under CPython on a Pentium box (866MHz P3 here):

>>> t1=clock(); x= 2L**10000; delay=clock()-t1; print delay
0.00165775490706
>>>

So that's how fast a good (but not optimal) pow runs:  the Linux number is
plausible.

The obvious way to implement m**n is to do O(n) multiplications.  A much
smarter way (which CPython implements) uses only O(log2(n)) multiplications.
For n==10000, n/log2(n) is very roughly about 1000.  That's in the right
ballpark for explaining a factor-of-2400 difference, which latter is very
hard to explain otherwise.

all-answers-are-in-the-source-code<wink>-ly y'rs  - tim