[Wien] Benchmark-Compilation options

Martin Hilgeman hilgeman at sgi.com
Fri Jun 16 14:46:28 CEST 2006


Dear Peter,

On Thu, 2006-06-15 at 08:06 +0200, Peter Blaha wrote:
> Just noticed, that your timing does not seem to be cortrect. Check 
> cputim.c; Some machines report the cputime in different units and you 
> should divide by 1024 instead of 100.

That's right, the current implementation in the various cputim.c files
is not portable. I'd advise it to change it like:

--- 00/SRC_lapw1/cputim.c       2002-12-09 06:04:22.000000000 -0800
+++ 01/SRC_lapw1/cputim.c       2006-03-01 03:12:15.158745600 -0800
@@ -10,7 +10,7 @@
       struct tms buffer;

       times(&buffer);
-      *dsec = (double)buffer.tms_utime/100.0;
+      *dsec = (double)buffer.tms_utime/sysconf(_SC_CLK_TCK);
 /*      fprintf(stderr,"inside cputim: %lf\n",dsec);*/
 }


> I've also heared from an Bull Itanium system and they gave very similar 
> timings as the Altix. So I guess these values should be reachable. Since 
> your "Hamilt-time" is ok (besides the factor 10), it is mainly the 
> gotoblas, which makes the difference. Also your report that you get the 
> same timings with/without goto tells me, that you are NOT using the 
> gotolib (or not a good version). I guess, the goto-libraries were changed 
> with version 1.0 and one has to compile them himself, so maybe older 
> versions (precompiled) were more efficient .... ???

I do not have first hand experience with self-compiled GOTO libraries,
only used the precompiled version when they were still available. I can
imagine that for a highly tuned BLAS implementation (like GOTO's), the
exact compiler version and options are extremely important.

> Always remember: The hamilt time is purely f90 + svml-lib code, hns is 
> mixed f90 + blas, and diag is purely blas-code. So compare the partial 
> times.

regards,

-Martin

-- 
Martin Hilgeman                    Chemical Applications Engineering
                                   Phone:  +31(0)30-6696885
SGI                                E-mail: hilgeman at sgi.com
The Netherlands                    URL:    http://www.sgi.com/go/chembio


More information about the Wien mailing list