Re: GMP performance on Opteron

[Bill Allombert <allomber@math.u-bordeaux.fr>]

On Mon, Oct 24, 2005 at 02:56:06AM -0400, Igor Schein wrote:
> Hi,
> 
> For the following command
> 
> setrand(1);rnfisnorminit(y^2+29*y+21,x^3-10*x^2-3*x-5);
> 
> GMP kernel gives a visible performace advantage on 64bit Sparc
> Solaris and 32 bit ix86 Linux over non-GMP kernel, but is slightly
> slower than non-GMP on Opteron.  Any idea why?

As far as I understand, GMP 4.x is not optimised on Opteron as much 
as on ix86 or sparc, while PARI include assembly code for Opteron, so it
might be that PARI is faster for small operands. 

You might want to try the patch I sent for lower GMP overhead to see
if it make a difference, see
<http://pari/archives/pari-dev-0508/msg00019.html>.

Cheers,
Bill.