don@cs.qmc.ac.uk (Don Beal) (09/28/88)
This article is being posted on behalf of Tony Pudner:
--------
From A. Pudner, British Aerospace.
Reference: "DLM - A powerful AI Computer for Embedded Expert Systems".
Article in "Future Generation Computer Systems" Vol 3, No 4, Dec 87.
The following consititute my reply to Richard O'Keefe's recent comments.
1) The above article describes 2 computers - an "Experimental DLM" and the
"DLM-1". The "Experimental DLM" became operational during December 1986
and was designed only to test the DLM architecture. For this reason the
memory areas were small and made from fast static RAM. The "DLM-1" was
designed on the basis of the experience gained during the experimental
phase. Unlike the experimental unit, the "DLM-1" is designed for "real"
programs and has a sophisticated memory architecture. Richard's comments
refer to the "Experimental DLM". However, it is the "DLM-1" that is the
real machine and comments should be focussed in this direction. Briefly
the parameters of the DLM-1 are:
Memory speed: 100ns (DRAM)
Memory working area: 9 Mbytes (upgradeable to 36 MB)
Program memory size: 2 Mbytes
Speed: 900 KLIPS (Deterministic concatenate)
2) The SUN-3 benchmarks (Quintus Prolog) are derived from the document
"ALVEY PROLOG BENCHMARKS" by S.Dickens and A.Bustany. This document
compares a number of Prolog benchmarks for different implementations
and computers. The document is very useful, but is now 2 years old. If
there are up-to-date figures for the SUN-4 and Quintus Prolog 2.4, I
would be very interested to receive them.
3) The parameter (Prolog Speed * Memory Speed) is only useful for
memory-bound computers. The DLM-1 is not memory bound, but for the
record its "figure of merit" is 90 mLi. This compares with 7.5 mLi for
the above SUN-3 implementation.
4) The benchmarks that I have done so far indicate that the DLM-1 is up to
100 times faster than the above SUN-3 implementation. The DLM-1 really
comes into its own on non-deterministic programs. In fact, Naive
Reverse, so far, gives the lowest performance ratio.
5) I am always interested in the opinions of other people on the DLM.
Since I only came across Richard's comments by accident (BAe is not linked
to usenet) I would be very grateful if any comments could be sent directly
to me. My address is British Aerospace, Downshire Way, Bracknell,
Berkshire, RG11 1QL, ENGLAND. Tel: +44 344 483222 Fax: +44 344 589000.
Many thanks, Tony Pudner.ok@quintus.uucp (Richard A. O'Keefe) (09/29/88)
In article <729@sequent.cs.qmc.ac.uk> don@cs.qmc.ac.uk (Don Beal) writes: >This article is being posted on behalf of Tony Pudner: >The following consititute my reply to Richard O'Keefe's recent comments. "recent" = 6 weeks ago. >2) The SUN-3 benchmarks (Quintus Prolog) are derived from the document >"ALVEY PROLOG BENCHMARKS" by S.Dickens and A.Bustany. This document >compares a number of Prolog benchmarks for different implementations >and computers. The document is very useful, but is now 2 years old. Prolog benchmarking is another topic, but I cannot agree that the document in question is "very useful". (With a month's notice I could go into detail about its flaws; at the moment it's swapped out of my memory.) >4) The benchmarks that I have done so far indicate that the DLM-1 is up to >100 times faster than the above SUN-3 implementation. The DLM-1 really >comes into its own on non-deterministic programs. In fact, Naive >Reverse, so far, gives the lowest performance ratio. A factor of about 25 is easy to account for based on gross features of the machine and implementation method. The additional factor of 4 is where the interest lies. It is particularly interesting to note that the WAM is optimised for determinate code. One of my complaints about the original article was that it said virtually nothing about the instruction set, data representation, or other aspects of programmer- visible architecture. Without knowing what that is like, the reader cannot tell how much of the performance is due to not using the WAM and how much of the performance is due to the "5 processing units and a sophisticated 2-stage program sequencer". >5) I am always interested in the opinions of other people on the DLM. >Since I only came across Richard's comments by accident (BAe is not linked >to usenet) I would be very grateful if any comments could be sent directly >to me. Sorry, but this once I'll impose on Don Beal's good nature. It would be most interesting if a bibliography of reports about the DLM-1 (and of papers which were influential in its design) were to be posted to comp.lang.prolog (or to the Prolog Digest).