implemented features (Re: Decision time....)

km b kmb810 at gmail.com
Mon Jun 4 04:28:22 PDT 2007


On 04 Jun 2007 10:46:20 GMT, Rahul Siddharthan <rsidd at online.fr> wrote:

Where is the generalisation?  I said "my own code", which is what I'm
interested in.  But it also appears to be true of GSL (the GNU
scientific library), at least the parts that I'm interested in.  I
don't think, even if my claim was made "generally", it would be at all
controversial.  Floating-point performance is much better in 64-bit
mode, and there are twice as many general-purpose registers available.
You're claiming that all 32-bit binaries are faster than 64-bit
binaries based on your specific experiments which is nothing but
generalization. And now you're saying floating point applications,
which is true due to the improved SSE instruction set in newer
generation CPUs from intel and amd.
In fact it is a tragedy that nearly all 64-bit computers out there
still run 32-bit Windows.  As a result you see benchmarks claiming
that the new Xeons cream the current AMD Opterons; but these are all
32-bit benchmarks, and say little about 64-bit performance.  For my
code (and some of my colleagues say the same), in 64-bit mode, the
Opterons still seem ahead.
>Have you measured against pure 32-bit environment vs
>pur 64-bit environment or just executing 32-bit binaries under 64-bit
>environment. Have you measured the effect on the cache in different
>environments (if you are using FreeBSD you can do so using PMC).
They are all 64-bit OSs (Linux or Solaris), but a 32-bit binary links
with 32-bit libraries.  I don't see why you think the cache usage
would be worse for a 32-bit binary.
Simple answer would be cache line size and how processor utilizes
cache for given virtual address.  You would see this effect more if
you were just on the edge of thrashing the cache in 32-bit mode,
64-bit mode would completely thrash.
Rahul

cheers
kmb
--
Something is wrong up on cloud # 9!




More information about the Kernel mailing list