Re: Good GMP performance (WAS Re: Problems linking with gmp...)

[Will Galway]

Torbjorn Granlund writes:
 > Kevin Ryde <address@hidden> writes:
 > 
 >   Usually 64-bit is faster, though I'm not sure by how much on sparc.
 > 
 > It should be about 2x faster for UltraSPARC.
 > That's still not very fast, of course.
 > 
 > If you care about speed, don't run GMP on any Sun hardware.
 > Athlon, Pentium 4, or Alpga 21264 are all much better for that.
 > 
 > (It would be possible to boost GMP speed with about 2x on
 > UltraSPARC 3, but that'd require quite a lot of work.)
 > 
 > -- 
 > Torbjörn

Thanks for the advice, Torbjörn.  Fortunately, I have a large (200
CPU) cluster of Athlon's to work with, and a few (32) Alphas of some
vintage (I don't know what vintage).  On the other hand, I also have
access to several (50 or so) UltraSPARCs, which are more likely to be
unused by others.

At some later date I'll probably send out a note asking for advice on
getting a higher-performance modular exponentiation routine, in the
hopes that I can do better than mpz_powm (which is what I'm using
now).  The speed at which I can do modular exponentiations is the
limiting factor in a computation I'm currently working on (searching
for pseudoprimes).  I'm hoping that I can do significantly better than
mpz_powm because for this application my needs are limited to
computing 2^k modulo n, where both k and n are (unsigned) 64-bit
integers.  Using mpz_powm, it looks like I'm getting roughly 10^5
modular exponentiations per second on a 1.2 GHz Athlon....

Regards.

--Will
( mailto:address@hidden http://www.cecm.sfu.ca/~wfgalway )