[comp.lang.fortran] Sun 4 Fortran benchmark results

rfinch@caldwr.water.ca.gov (Ralph Finch) (06/12/90)

I've always wanted to compare our 4/260, with the old Weitek floating
point unit, with our 4/330, which runs 50% (I think) faster by clock speed
and has the new TI fpu.  A benchmark run on a Cray Y-MP of one of our
models spurred me into benchmarking the Suns; here are the results.

The model uses floating point fairly much, but does not vectorize well at
all on the Cray.  Some I/O, probably not significant.  SunOS 4.0.3,
Fortran 1.2.  The compile types in the table below are as follows:

dbg:	-g
opt:	-O
optmax:	-O /usr/lib/f77/nosqrt.il -dalign	(4/260)
optmax:	-O /usr/lib/f77/libm.il -dalign		(4/330)
optmax-oth:	compile as optmax on one machine, run on the other

The optmax libraries are as described on pp. 4-5 of the READ THIS FIRST
for the Fortran manual.

Results here are run times relative to the fastest time, which is the
4/330 with optmax.  Original run times, in seconds, were had by adding the
'user' and 'sys' times output by the /bin/time command.  In all cases
except the 4/260:optmax-oth one the sys time was a very small fraction of
the user time.  In the other case, the sys time was almost 18 times as
large as the user time; I suppose this is because it was doing some or all
of the floating calcs in software.

Machine                           Compile
                                    Type
                        dbg     opt     optmax  optmax-oth

4/260                   3.68    2.42    2.41    48.8
4/330                   2.03    1.17    1.00    1.18

Conclusion: It's worthwhile using the inline libraries if you have the new
TI fpu, but if you have mixed machines on the net, beware of running that
executable on machines with the Weitek fpu.  I keep the executables
separated with a _W or _T filename extension, generated automatically in
the makefile with the fpu checker program I posted here a few weeks ago.

The Cray ran the program about 14 times as fast as the 4/260 with optmax,
which equates to about 6 times as fast compared to the 4/330.  This is not
a poor reflection on the Cray, rather on the model, which doesn't
vectorize much if any.

Ralph Finch			916-445-0088
rfinch@water.ca.gov		...ucbvax!ucdavis!caldwr!rfinch
Any opinions expressed are my own; they do not represent the DWR