On Thu, 14 Aug 2003, Henry Schaffer wrote:
> > ... Now we are looking for Benchmarks to test the system ...
>
> John mentioned the desirabililty of benchmarking with something
> relevant to your workload - and I'll second that. Also it is nice to
> start with something relatively small and simple to make it easier to
> work with and to analyze the results.
>
> My interest is in scientific computation, and I wanted to compare a
> number of systems for "typical" floating point computation. (Yes, I
> know that this isn't the strong point of most of the systems we discuss
> here.) I talked with a expert, and came up with a very simple C program
> which does a bunch of matrix multiplications - and which has data in
> them so the compiler can't reduce ("optimize out") the computation.
> Also the different sizes used may show if system limitations are
> reached. I've run it on a variety of Linux and Solaris systems.
> --
I can't resist a benchmark;-)
These are the bottom lines from my various runs. They took 73 minutes
all-up. Here's the script:
[EMAIL PROTECTED]:~$ cat BM
#!/bin/bash
set -xe
for o in s 1 2 3
do gcc-2.95 source/b..c -O${o} -o t
./t >gcc-2.95-${o}
gcc-3.0 source/b..c -O${o} -o t
./t >gcc-3.0-${o}
done
[EMAIL PROTECTED]:~$
The numbers:
[EMAIL PROTECTED]:~$ tail -1 $(dir -1 gcc-* | sort)
==> gcc-2.95-1 <==
1000 106.92
==> gcc-2.95-2 <==
1000 94.40
==> gcc-2.95-3 <==
1000 93.65
==> gcc-2.95-s <==
1000 97.60
==> gcc-3.0-1 <==
1000 105.87
==> gcc-3.0-2 <==
1000 105.54
==> gcc-3.0-3 <==
1000 104.61
==> gcc-3.0-s <==
1000 104.64
[EMAIL PROTECTED]:~$
I'm intrigued that the older compiler apparently produces better code.
Now to try CPU-specific optimisations, and other CPUs.
--
Cheers
John.
Join the "Linux Support by Small Businesses" list at
http://mail.computerdatasafe.com.au/mailman/listinfo/lssb
Copyright John Summerfield. Reproduction prohibited.
