Remarks from Dr. Taeschner
From: Michael Taeschner
To: templon@studbolt.physast.uga.edu
Subject: results on i486/100 for silly_bench (brem)
Date: Tue, 9 Dec 1997 18:14:59 +0100 (MEZ)
Dear Prof. Templon,
found your webpage with the benchmark and tried it on my i486
running linux-2.0.30 (SuSE 4.4.1). This PC has 256 KB L2 cache
and 32 MB FPM DRAM. CPU is Intel 486DX4 100 MHz.
Here are my results:
xterm console
Absoft binary: 1245s 1108s
PG binary: 1366s
F2c binary: 1527s
g77-0.5.21: 1446s (1) 1260s (1)
egcs-2.90.14: 1440s (1) 1266s (1)
(1) compile options: -O2 -fomit-frame-pointer -finline-functions
-funroll-all-loops
Remarks:
1. For my runs I used g77-0.5.21 based on gcc-2.7.2.3.f.1 and egcs-2.90.14
(snapshot as of 971114, should give similar results as release
1.0). Egcs seems to bring no improvements for 486, but it is said to
have better optimization for 586 and 686. I compiled the sources
suffixed f2c, since g77 is due to historical reasons related to f2c.
2. The original source code would be somewhat more portable if you used
print*,
or even better
write(*,*)
instead of:
type*,
g77 complains this is an unsupported VXT extension and indeed, I've
never seen it before.
[ remark from JAT -- indeed, this is a VAX extension. I will remove it
soon. ]
[ ...]
3. ldd brem_absof shows it is only dynamically linked to libc.so.5 which
resolves to /lib/libc.so.5.4.7 on my system),
where as the brem_pg uses:
libm.so.5 => /lib/libm.so.5.0.6
libc.so.5 => /lib/libc.so.5.4.7
and brem_f2c in addition
libf2c.so.0 => /usr/lib/libf2c.so.0.11
which is much older than the static version as of g77-0.5.21. One
might suspect, that at least absoft benefits from optimized libm and
runtime libraries. Possibly this could be proved by gprof or use of
the equivalent tools of the commercial compilers (by comparing times
actually spend in your code and times needed for executing library functions).
The other day someone pointed to me out, that he gets much better timings on
code compiled by egcs using a specially patched libc and libc compiled by
Pentium gcc than using plain gcc.
[ ... ]
5. For me runnung in console mode (without X11 in the background) gave
far better results. Did you really get better results when running
under X11 than running on console?
Jeffrey Templon
Last modified: Tue Dec 9 23:22:51 EST 1997