mprime on SunOS/sparc?

[rudi_m]

best-FFT timings appear in the mlucas.cfg file that results from the self-tests

Ah ok,
I will try that but currently I run Glucas which is about 3 times faster as the my first Mlucas binary (maybe the "configure" script in Glucas just made a better job as my "cc -o Mlucas *.c -lm" for compiling Mlucas ;).

If you have multiple CPUs, the reported utilization may be in terms of the overall system utilization, e.g. with 4 CPUs a process running on only one will get at best 25%. You have 24 CPUs, but each runs at only 150MHz, did I read that right?

Currently it seems to be 50 CPUs each 1350MHZ with 80G RAM
but Ive heard that its a dynamic system which is bigger during day and smaller at nights.

George already has optimized code for AMD in Prime95/mprime, and AFAIK the Intel C compiler (far better than gcc for x86-style CPUs) isn't available for AMD, so there's little point in using anything other than mprime on AMD at present.

Yep, of course - I was just comparing to learn about Mlucas.

Do a "which cc" and/or "man cc" to find out - but I'm 99.9% certain that it's gcc, which in my experience absolutely sucks at producing optimized code for Sparc. For decent performance, you've got to use the SunPro (native C) compiler

well the cc in $PATH seems to be SUNWspro
Now I tried also gcc 3.4.1 and you are right it was 20% slower!
But with gcc -O3 It was double as fast!!
Sou would be really interesting to checkout SUNWspro Optflags too!

- if anyone wants to give me guest access to a system with that compiler, I'll be more than happy to do some building and tuning and post a prebuilt executable to the Mlucas page.

I cant give you login to that machine because its just a friend's login
But I watch through my own accounts if I find another Sun/Sparc.

Or you could even send me bash scripts with compile runs and benchmarking etc. which I could run for you and send you the results back.

[rudi_m]

- if anyone wants to give me guest access to a system with that compiler, I'll be more than happy to do some building and tuning and post a prebuilt executable to the Mlucas page.

OK, Ive sent you login to a similar machine- watch your PM.

[ewmayer]

OK, I've played around with various build flags on the Ultra-3 system Rudi got me a login for, and got a preliminary binary built. The main thing I discovered while doing the builds was that the #define sequence I had previously been using to inline prefetch macros on Sparc with the SunPro C compiler did not work with the more-recent versions of the compiler - it seems Sun changed the associated predefines. Once I'd fixed that up, timings improved by 20-25% across the board, though they're still not what I'd hope to see from this type of processor (compared to, say, timings on the Alpha ev6, which has similar floating-point capabilities). Here is what I'm getting at present on a 900 MHz Ultra-3:

FFT length (K)     Timing
576               sec/iter =    0.118
640               sec/iter =    0.135
704               sec/iter =    0.160
768               sec/iter =    0.152
832               sec/iter =    0.182
896               sec/iter =    0.199
960               sec/iter =    0.198
1024              sec/iter =    0.206
1152              sec/iter =    0.249
1280              sec/iter =    0.293
1408              sec/iter =    0.365
1536              sec/iter =    0.366
1664              sec/iter =    0.431
1792              sec/iter =    0.475
1920              sec/iter =    0.485
2048              sec/iter =    0.500

These are at least 1.5x worse per-cycle than I get on Alpha - I notice that the compiler on the system I'm using is not the latest (it's v9), so I'm going to look around for a system with a v10 compiler installed and see if the performance improves any.

Rudi, could you do some Glucas timing runs on this same system (so we get a fair comparison) and let me know what sorts of timings you get at these kinds of FFT lengths?

[rudi_m]

Rudi, could you do some Glucas timing runs on this same system (so we get a fair comparison) and let me know what sorts of timings you get at these kinds of FFT lengths?

Sry, I got less time for playing around lately.
But you can use that account as long as you want for doing further tests.

[ewmayer]

Originally Posted by rudi_m

Sry, I got less time for playing around lately.
But you can use that account as long as you want for doing further tests.

...except that I've barely got enoughh time to do experimental builds of my own code, much less someone else's. You said you've built glucas on the sparc - could you point me to a binary?

[rudi_m]

ok, I compiled glucas on that machine using the sun compiler and gcc.
You will find glucas-sun and glucas-gcc in $PATH there.
First tests showed that the sun compiled binary is a bit slower but maybe the configure script just handles the gcc in a better way.

[ewmayer]

FYI, I posted a copy of my current Sparc binary (the one giving the above table of timings) to

ftp://hogranch.com/pub/mayer/bin/Mlucas_sparc_v9.gz