home
***
CD-ROM
|
disk
|
FTP
|
other
***
search
/
Simtel MSDOS 1992 December
/
simtel1292_SIMTEL_1292_Walnut_Creek.iso
/
msdos
/
microcrn
/
issue_50.arc
/
UNIXBNCH.ARC
/
UNIXBNCH.DOC
Wrap
Text File
|
1989-09-21
|
8KB
|
174 lines
Editor's note: We didn't have room to print much of Bob's UNIX
benchmark info in issue #50 of Micro Cornucopia, so here goes:
Benchmarking
One goes about this differently than with a DOS system. DOS
benchmarks are produced mainly by competing magazines and
commercial concerns. Fortunately, Unix benchmarks were produced
mainly by academics who did quite a bit of research. The "128k
no-op" and other such garbage are notably absent.
A systems approach is used. There are three basic areas:
execution of general code, mathematical computation, and file
system performance. Factors such as multiuser loading will not
differ from one product to the other because every system except
XENIX uses the same scheduler routine.
There are three most esteemed tests. The Dhrystone was
written by Reinhold Weicker and Rick Richardson. It is the result
of a statistical analysis of program statement frequency. Since
it is almost always compiled from C on the system you are
testing, it takes into account compiler performance as well as
processor speed.
The Whetstone, by Curn, is usually executed on systems with
a math coprocessor. No 386 Unix systems currently use inline
coprocessor instructions. Thus this test measures the speed of
the kernel based coprocessor routines, for which there is
surprisingly wide variation. Since the test machine does not have
a 387, I have not performed this test. An Inboard 386/AT so
equipped delivers 800k/sec Whetstones identically for all the
Unix ports, while the result for XENIX is 550k/sec.
The Iostone, written by Arvin Park and Richard J. Lipton, is
referred to as a "synthetic file system benchmark." It operates
on a 4 megabyte file with reads and writes which were generated
by examining the statistics of numerous installations. Along the
way the author discovered an apparently universal law: reads are
twice as numerous as writes. The block size varies from 256 bytes
to 65K, according to a biased random number generator.
I don't feel completely comfortable with synthetic
benchmarks because there is always the possibility of a flaw. And
it's difficult to integrate the numbers into an estimate of how
fast an application will actually run. Since I'm a programmer, I
devised a test which is an actual compile and link sequence. The
code I used is a public domain logic synthesis program called
ESPRESSO, written at the University of California at Berkley.
What it actually does is of little interst to readers. The 100k
of source code is can be compiled on virtually any system.
Since a slow display system can be an irritant, I dump all
the source to the screen, timing how long that takes. For timing
I use the "time" utility, as in the following lines:
time make<CR>
time cat *.c *.h
The "make" utility compiles ESPRESSO. The "cat" utility is
similar to "type" under DOS.
One pleasant surprise is that the 386 system which you put
together out of clone parts is about five times faster than a VAX
780, or about twice as fast as a VAX 8600. It has nothing to fear
from SUN 3 workstations either. Although RISC chips are faster,
your PC based workstation would be either as fast or faster than
the large base of installed workstations.
If you are familiar with a V.3.0 port, be aware that V.3.2
is much faster. For example, the Interactive Systems compiler and
file system are twice as fast as their 3.0 counterparts, while
the file system is five times as fast as Microport Unix.
The benchmarks were performed on a Bell Tech (now part of
Intel) MPE386-25. With addressing for 1/4 gigabyte of memory,
it's an ideal Unix platform.
Benchmark Results
Operating System
XENIX 386/ix Intel ESIX
Dhrystone2 NO REG 7923.2 7698 7698.2
Dhrystone REG 7874.0 7942 7803 7942.8
Iostone 2380 7142 5633 5402
C&L 2:50.3 2:26.7 2:43 2:49.4
L 2.8s 2.1s 2.6 2.8
cat to screen 18.7s 1:12 1:16 1:08
C&L = ESPRESSO compile and link
L = link only
cat = send all ESPRESSO source files to the screen
"REG" refers to the use of register declarations in the code. The
number of disk buffers was set to 500. However, the iostone
results are actually quite insensitive to the number of buffer
provided they do not approach the size of the 4 megabyte test
file.
The comparison figures allow you to compare the 386 with
other processors. The results are quite favorable. Readers with a
modest budget will be encouraged by the results of machine #2 -
it's an INBOARD 386 accelarator board, which I am quite satisfied
with as a personal system.
Dhrystone Comparison Figures
Manufacturer/Model Dhrystones
Sun 3/60 (20 mhz 68020) 4273
Sun 3/280 (25 mhz 68020, cache) 6329
DEC VAX 11/785 2090
DEC VAX 8600 3866
IBM PC/AT, large model 696
ESPRESSO Compile and Link Comparison Figures
Operating System | Machine |Disk buffers|________Time____________
(same versions) | | | Total | Link Only
-----------------------------------------------------------------
XENIX | 2 | 537 | 4:04 | 5.1
386/ix (V.3.0) | 1 | 500 | 4:34 | 5.3
Microport | 1 | 1100 | 5:02 | _
Sun OS | 3 | _ | 4:04 | _
Sun OS | 4 | _ | 3:54 | _
Bell Tech Unix | 2 | 200 | 4:00 | 6.1
Bell Tech Unix | 2 | 500 | 3:55 | 3.7
Bell Tech Unix | MPE/25 | 200 | 2:41 | 5.4
Bell Tech Unix | MPE/25 | 500 | 2:43 | 2.6
Iostone Benchmark Comparison Figures
Operating System | Version | Machine |Disk buffers|Iostones
| |(kilobytes) |
-----------------------------------------------------------------
XENIX | 2.3 | 2 | 537 | 2279
386/ix (V.3.0) | 1.06 | 1 | 500 | 1990
Microport | 2.2 | 1 | 1100 | 1049
Sun OS | 3 | 3 | _ | 1834
Sun OS | 3 | 4 | _ | 3669
Bell Tech Unix | V.3.2 | 2 | 200 | 4597
Bell Tech Unix | V.3.2 | 2 | 500 | 5263
Bell Tech Unix | V.3.2 | MPE/25 | 200 | 5280
Bell Tech Unix | V.3.2 | MPE/25 | 500 | 5633
Hardware Configurations Used
Machine #1: Inboard 386/AT at 16 mhz, 4meg ram, WD1003-WA2 MFM
controller, 2:1 interleave, two MAXTOR XT-1085 hard disks.
Machine #2: Inboard 386/AT at 16 mhz, 4meg ram, Western Digital
1006V-SR2 RLL controller, 1:1 interleave, two MAXTOR XT-1085 hard
disks.
Machine #3: Sun 3/60, 8 megabytes of RAM, local disk, 16 mhz
clock speed.
Machine #4: Sun 3/280 file server. 25 mhz cached 68020, 900
megabyte Fujitsu Eagle Hard disk, SMD controller.
MPE/25: Bell Tech 25 mhz, based on Intel 302. CDC 380 meg ESDI
drive, 3.8 megabytes of RAM, ESDI controller compatible with
WD1007.