Brian Beesley writes:

<< I loaded Mlucas_2.7y onto a Ultra 10 (128 MB, Ultra IIi @ 300 MHz, 
Solaris 2.6) & have the following timings. The comparison 
MacLucasUNIX timings were generated using gcc with options 
-O6 -mpcu=v9 -Wa,-xarch=v8plusa, using the same method I used for the 
timings I sent you for my Alpha system at the weekend. These are 
about 20% slower than the "production" version compiled with the Sun 
C compiler (binary provided by Simon Burge; I found the "Ultra 5" 
version was actually the fastest on my system).

I was unable to benchmark MacLucasUNIX at 4096K FFT run length due to 
insufficient system memory - might _just_ manage with 192 MB.

Format:
FFT size, Mlucas, MacLucas

64K, 0.040, 0.049
80K, 0.058, n/a
96K, 0.079, n/a
112K, 0.097, n/a
128K, 0.116, 0.142
160K, 0.155, n/a
192K, 0.187, n/a
228K, 0.228, n/a
256K, 0.254, 0.306 (Production MacLucasUNIX = 0.228)
320K, 0.328, n/a
384K, 0.400, n/a
448K, 0.481, n/a
512K, 0.542, 0.629
640K, 0.718, n/a
768K, 0.841, n/a
896K, 1.041, n/a
1024K, 1.120, 1.333
1280K, 1.544, n/a
1536K, 1.826, n/a
1792K, 2.063, n/a
2048K, 2.320, 2.781
2560K, 3.152, n/a
3072K, 3.864, n/a
3584K, 4.591, n/a
4096K, 5.209, n/a

You will see that, even allowing for the production version of MLU 
being faster than the instrumented version, Mlucas is about the same 
speed or faster, except for any exponents for which both programs use 
the same FFT size. These are fairly rare. >>

Hi, Brian, and thanks for the Ultra 10 timings and other comments.
Regarding the former, could you provide me with all the 64K-2048K
MLU timings using the "production" version of MLU? (Not just 256K).
Also, I'd appreciate the cache sizes of your Ultra10/300, if you
know them.

>There is a documentation bug relating to the environment variable 
>LD_LIBRARY_PATH for the Solaris version. If you have no setting for 
>this variable & just point it at the directory containing the 
>libraries distributed, you will find that most of the Sun utilities 
>stop working! You will probably want to do something like
>
>LD_LIBRARY_PATH=/usr/lib:$HOME/lib
>export LD_LIBRARY_PATH
>
>so that the default system libraries are searched first. Mlucas 
>doesn't mind, and the other utilities seem to prefer things this way!


When Bill Rea (who put together the SPARC binaries) told me the
LD_LIBRARY_PATH environment variable was named the same on Sun as on
Alpha, I assumed all else would work as it does on Alpha Unix, too.
I'll correct the documentation as you suggest.

>1) Can I use "Prime95" format worktodo.ini files (i.e. 
>Test=exponent,depth or DoubleCheck=exponent,depth) or must I just 
>have the exponents, one per line?

The latter. We won't need the rest of the Prime95 format until Peter
Montgomery's Mfactor code is included as a module. Mlucas doesn't
differentiate between first-time tests and double checks, but that
is also moot for now.

>2) If I want to add/remove/resequence assignments in worktodo.ini, 
>must I do this with the program stopped, or does it reopen the file 

You can change anything but the first line (currently active exponent)
while the program is running. When the current exponent finishes, the
program opens the worktodo.ini file, deletes the first entry, and starts
on the next exponent. Now that you mention it, it would probably be a
good idea to have the program check the first entry of the worktodo.ini
file against the just-completed exponent, to make sure the user didn't
inadvertently put a new exponent on line 1 once the old one was started.
I'll put that on my to-do list.

Keep those timings/suggestions/bugreports coming,
-Ernst

_________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
Mersenne Prime FAQ      -- http://www.tasam.com/~lrwiman/FAQ-mers

Reply via email to