On 3 Nov 99, at 9:50, Bill Rea wrote:
> If there are people with SPARC systems with large external caches
> thinking about using Ernst Mayer's Mlucas for LL testing then
> make the switch.
I actually switched my Ultra 10 to Mlucas (using the binary you
supplied for Solaris 2.6) a couple of weeks back. I don't know what
the cache size on this system is, but I doubt it's as big as 4MB.
The point is that I found (a) MLU at FFT size N is about the same
speed as Mlucas at FFT size 7N/8, but (b) this is offset to a large
extent by the fact that Mlucas runs larger exponents than MLU with
the same (power of 2) FFT size.
e.g. on exponents around 3.7 million the iteration time for Mlucas is
0.185 using 192K FFT size. MLU would need a 256K FFT size and would
run at 0.228 sec/iter. (Ultra IIi @ 300 MHz)
Of course, there will be _some_ exponents where MLU is a bit faster
on this particular system - so I'll keep MLU around.
Bear in mind that Mlucas is still a bit "experimental". One problem
has come to light - if you're using the PrimeNet Manual Testing page
to submit results, you _must_ remove the space preceding the M at the
beginning of the result line, else PrimeNet won't accept the result.
Actually it says it has got the result but doesn't remove the
assignment from the current assignments report or add it to the
completed assignments report unless the leading space is removed.
(This seems to be a program bug, I see the extra space in the source
code so I'd assume the problem afflicts Mlucas 2.7y on all platforms.
Certainly the binary executable I built for Alpha systems running
linux is also afflicted.)
> MLU is unusual in that it runs significantly slower in 64-bit mode.
Very odd. Any idea why this might be? Compiler producing broken
code???
Regards
Brian Beesley
_________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
