moskowit@paul.rutgers.edu (Len Moskowitz) (10/10/90)
Sun 3/80 and Tadpole TP881V (88000 RISC)
Benchmarks for CLIPS Production Systems
DISCUSSION
Back in June and July I posted a few queries about small scale parallel
architectures (less than 64 computing nodes) for hosting production system
languages (e.g., CMU's OPS5 and NASA's CLIPS). Since that time we've narrowed
our choice of microprocessor to the Inmos Transputers (the T800 and T801) and
the Motorola 88000 RISC. One of our constraints is that the final system must
occupy no more than three 6U VME slots. Both Inmos and Transtech have 6U
boards (the B014 and the MCP500 respectively) that accommodate eight size-1
transputer modules ("TRAMS"). Tadpole Technologies has a three board set based
on Motorola's Hypermodule that hosts eight 88100 CPUs and sixteen 88200
cache/MMUs.
We decided to initially benchmark single processor configurations of the
Transputer and 88000 versus our Sun development system and embedded target
hardware (the Motorola MVME-147). I reported on the results of the Transputer
benchmarks in a previous post. This report presents the results of the 88000
benchmarks and compares them with the performance of the Transputer, Sun 3/80,
and Motorola MVME-147 board.
Tadpole Technology loaned us a single processor version of their TP881V 88000
VME development system for evaluation. It consisted of three boards: a
Processor module holding a single 20 megaHertz 88100 CPU, 2 88200 Cache Memory
Management Units, and 8 megabytes of memory (it can hold up to 4 25 megaHertz
88100s, 8 88200s, and up to 32 megabytes of memory); a Base module providing a
VME interface, disk i/o, network i/o, and console i/o; and a VSB module that in
our system held no memory at all and was unused. The three boards resided in a
stand-alone VME chassis. A simple CRT terminal acted as console. The Tadpole
system connected to our Sun network via NFS.
It took less than an hour to set up the Tadpole system and to bring it to our
Sun network. Porting of CLIPS 4.3 and our application to the TP881V using
Tadpole's TP-CDS C compiler took less than 6 hours. We were able to use the
benchmarks used for the Transputer, the Sun, and the Motorola board with no
modifications. We required no technical assistance to get the benchmark
running. This is a tribute to both CLIPS' portability and the efficiency of
the Tadpole tools.
Our application programs are typically developed on Suns and ported to a
Motorola MVME-147 single board computer that is embedded in our products. The
-147 uses a 25 megaHertz 68030 and 68882. The Sun 3/80 uses a 20 megaHertz
68030 and 68882 floating point coprocessor. The TP881V we evaluated uses a 20
megaHertz 88100. The -147 runs VxWorks 4.0, a real time UNIX kernel. The Suns
run SunOS 4.1. The Tadpole TP881V runs TP-IX/88K, a multi-processor UNIX
operating system for the 88000, based on kernel technology licensed from Data
General. The -147 uses 70 nanosecond, 1 wait state memory. We don't know the
memory parameters for the Sun or the Tadpole. Both the Sun and the -147 run
code generated by the same Sun C compiler, albeit linked with different
operating system service libraries. The Tadpole used code generated by the
Tadpole TP-CDS C compiler. All processes on the -147 other than the benchmark
were suspended (these suspended processes occupied no more than one percent of
the processing capacity of the -147). The Sun's usual processes and their
priorities weren't interfered with in any way except that remote logins were
disabled. The Tadpole system was in multi-user mode but ran only the benchmark
program; no other processes were active.
The benchmark, an expert system for waveform analysis consisting of 71 rules
averaging 4 condition elements and 4 actions, performed no output to the
display and required no disk access. All data required for execution (the
working memory elements) were preloaded into working memory. The rules were
pre-compiled.
RESULTS
The timing results reported below are the average ratios for 20 different data
sets. Timing ratios (TP881V versus Sun 3/80) were very constant for the 20
data sets, varying less than 9 percent from the mean. Timings were measured
from the invocation of the production system until rule firing ceased.
The 20 rule sets took between 83.98 and 201.09 seconds to run on the Sun 3/80.
Timing ratios varied from a low of 9.15 to a high of 10.18.
The Tadpole TP881V with a single 20 megaHertz 88100 CPU was on average 9.93
times faster than the Sun 3/80. As reported in the previous benchmark, the
B404-5A Transputer was 2.24 times faster than the Sun 3/80 making the Tadpole
TP881V roughly 5 times faster than the B404-5A Transputer. However, recent
timings performed by INMOS using our benchmarks on the fastest transputer
module available to them today (a B428-5) show that it is 4 times faster than
the Sun 3/80. This leads us to believe that the 20 megaHertz TP881V is 2.45
times faster than the B428-5 Transputer.
In previous benchmarks the -147 board was uniformly about fifty percent faster
than the Sun 3/80, making the Tadpole TP881V roughly 7 times faster than the
-147.
We expect that the readily available 25 megaHertz version of the TP881V will
significantly better these figures.
CREDITS
The benchmarks were designed and run by Dr. Elliott Greene at Bendix Test
Systems Division, Teterboro, New Jersey, 07608, USA.
Mike Roche at Tadpole's Long Island (New York) office, and Alice Stifter at the
Austin, Texas sales office were both very helpful. We thank them both for the
loan of the evaluation system and their support.