.\"	@(#)nroff.in	1.1 97/01/03 Connectathon Testsuite
.DA
.ds RF Company Confidential
.ds LF Sun Microsystems
.nr PS 12
.nr VS 14
.ps 12
.vs 14
.TL
Fortran Benchmarks from General Electric
.AU
Evan Adams
.LP
Walter Sawka from the New York office sent a memo
on June 21, 1983 detailing fortran performance.
It included the source and results to three fortran
benchmarks.
.LP
I ran the benchmarks on the following configurations:
.TS
center;
l l l l.
CPU	Unix	Fortran	Memory
.sp 4p
Sun 1	Version 7	SVS	1 Meg
Sun 1.5	Sun 0.4	Berkeley	2 Meg
Sun 2	Sun 0.3	Berkeley	2 Meg
Sun 2	Sun 0.3	Berkeley/Optimizer	2 Meg
VAX 750	4.1cBSD	Berkeley/Optimizer	4 Meg
.TE
.LP
The benchmarks are:
.IP 1)
Matrix inversion of double precision values.
.IP 2)
A bubble sort of integers.
.IP 3)
A prime number generator.
.LP
.TS
center;
 c|c|c s|
 c|c|c s|
 c|c|c s|
 c|c|c s|
|l|n|n|n|.
	_	_	_	_	_
.sp 4p
	Matrix Inversion	Bubble Sort	Prime Numbers
.sp 4p
	_	_	_	_	_
.sp 4p
	Dimension	Number of Elements	Highest N
.sp 4p
_	_	_	_	_	_
.sp 4p
Configuration	50	1000	2000	10000	50000
.sp 4p
=
.sp 3p
Version 7	91.7	23.1	88.9	24.3	180.1
.sp 3p
_
.sp 3p
Sun 1.5	74.9	29.0	112.7	49.0/28.5\(dg	290.3/188.6\(dg
.sp 3p
_
.sp 3p
Sun 2	60.5	22.8	88.0	39.1/22.7\(dg	231.4/150.6\(dg
.sp 3p
_
.sp 3p
Sun 2 (OPT)	55.9	15.8	59.9	38.3/22.7\(dg	227.6/150.6\(dg
.sp 3p
_
.sp 3p
Apollo 68000	61	30	122	19	164
.sp 3p
_
.sp 3p
Apollo DN300	51	25	94	16	130
.sp 3p
_
.sp 3p
Apollo (PE)	28	22	88	14	127
.sp 3p
_
.sp 3p
VAX (OPT)	19.3	12.1	46.4	14.5	87.2
.sp 3p
_
.TE
.LP
\(dg using a single precision square root function
.LP
All times are in seconds.
The times for the Apollo machines are from Walt's memo.
His memo claims the times are elapsed (wall clock) time.
This seems silly since most of the programs prompt for some information.
The times for V7 and 4.2 are user time as given by the /bin/time command.
The Apollo is a 68000 at 10Mhz,
the Apollo DN300 is a 68010 at 10MHz, and
the Apollo (PE) is a 68000 at 10Mhz with the performance enhancement option.
.LP
The Apollo numbers for the prime number generator 
with a highest N of 10000 look suspect.
I find it difficult to believe that the Apollo (PE)
beat the VAX in this benchmark while it was
consistently beaten badly in the other benchmarks.
.LP
Berkeley fortran fared poorly in the prime number generator.
This program makes many calls to the square root function.
The square root function is part of the math library and is written
in C.
To take the square root of a single precision number
it is converted to doubled precision and passed to sqrt().
A double precision square root is calculated and returned where
it is promptly converted to single precision.
Approximately 75% of the execution time was spent in sqrt() and its
descendants.
.LP
I wrote a single precision sqrt() routine in 
fortran and the prime number generator ran 43% faster on 
the Sun 2.
.LP
The fortran optimizer mainly improves array references within 
inner loops.
The bubble sort improves by 30%; the prime number generator
does not change at all.
.LP
It should also be noted that the system times on the 4.2 systems were
2 to 7 times greater than the system times for the version 7 system.
On version 7, the system time averaged 2.4% of the user time.
On the Sun 2, the system time averaged 8.2% of the user time.
.SH
Conclusions
.LP
The same object code ran on the Sun 1.5 and the Sun 2.
The Sun 2 execution times averaged 20.6% less than the Sun 1.5.
There was speculation that the Sun 2 performance would improve
by as much as 30% as compared to Sun 1.5.
Some people may be disappointed by these numbers.
.LP
The Sun 2 with the fortran optimizer beats the best Apollo
time by 28% for the sorting program.
I believe that the Apollo performance enhancement option includes
hardware floating point.
Still, the Sun 2 is 9% slower than the Apollo DN300 on the matrix
inversion program and a factor of two slower on the prime number
generator.
Taking into account the double precision/single precision problem
for the prime number generator, the Sun 2 is 13% slower
than the Apollo DN300.
.LP
The fortran optimizer is still in a developement stage and
is not very reliable yet.
It is not ready for release!
.SH
Future
.LP
The benchmark numbers give us a reflection of where
we are at today.
The ultimate goal is
.IP 1)
A Sun 2 CPU,
.IP 2)
Unix 4.2BSD (not 4.1c),
.IP 3)
the SKY floating point board,
.IP 4)
the fortran optimizer,
.IP 5)
a single precision and double precision math library, and
.IP 6)
an improved fortran I/O library.
.LP
Development is in progress for everything but the libraries.
It is conjecture, but with each of these components in place
we should beat Apollo consistently.