mersenneforum.org

mersenneforum.org (https://www.mersenneforum.org/index.php)
-   Mlucas (https://www.mersenneforum.org/forumdisplay.php?f=118)
-   -   mlucas on sun (https://www.mersenneforum.org/showthread.php?t=1830)

ewmayer 2004-01-07 15:24

Thanks for the timings, delta_t. Looks like RTP is giving from 0-15% speedup for the FFT lengths of current interest to GIMPS (roughly from 512-2048K).

I've PMed some e-mails from Bill Rea regarding his builds of a little over a year ago using a then-new version of the SunPro C compiler.

One other note re. the mlucas.cfg file entries: you should remove (at least from the working-directory version) any FFT lengths for which the best runtime is still worse than that for the next-higher FFT length. This typically happens for FFT lengths which involve a large-ish odd first-pass radix, which will be relatively inefficient compared to an even or power-of-two radix. For instance in your timings we see that 960K (which uses a radix-15 front-end DFT pass) is slightly slower than 1024K, which uses all power-of-two passes. Removing the 960 line from the file will cause the code to not use that length in production-test mode, but obviously you could srtill use it for manual timing tests.

Finally, since it looks like you're running on a multi-CPU system, it would be interesting to see how the timings are affected by memory bus contention when there are multiple instances of the program running.

sunosx 2007-10-02 22:13

I tried compiling the source archive suggested by ewmayer on my [URL="http://www.anysystem.com/workgroup-servers-sunfire-4800.html"]Sun Fire 4800[/URL] running Solaris 9. Even though the installation was a bit troublesome, but it runs smooth for now. Thanks

ewmayer 2007-10-03 21:36

[QUOTE=sunosx;115559]I tried compiling the source archive suggested by ewmayer on my [URL="http://www.anysystem.com/workgroup-servers-sunfire-4800.html"]Sun Fire 4800[/URL] running Solaris 9. Even though the installation was a bit troublesome, but it runs smooth for now. Thanks[/QUOTE]

Could you be a bit more specific about "troublesome"?

We are trying - in the little spare time our work-for-pay leaves us - to get a beta 3.0 release ready by year's end. Tom Duell of Sun has built the current in-dev codebase for this on various systems including some based on the new Sparc 6 chips, and gotten good performance on the latter. Currently working on specific optimizations targeting Solaris-on-AMD64 systems. We wish we could work full-time on this, but that luxury is presently not available to us.

rgiltrap 2007-10-04 05:45

[QUOTE=ewmayer;115650] We are trying - in the little spare time our work-for-pay leaves us - to get a beta 3.0 release ready by year's end. [/QUOTE]

The 3.0 beta gives pretty good levels of parallelization which should be nice on the SF4800.


All times are UTC. The time now is 02:51.

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2021, Jelsoft Enterprises Ltd.