perm filename TMP[1,LMM] blob
sn#617858 filedate 1981-10-13 generic text, type C, neo UTF8
COMMENT ⊗ VALID 00010 PAGES
C REC PAGE DESCRIPTION
C00001 00001
C00003 00002 The research computing in the Computer Science Department is dominated
C00021 00003
C00023 00004 Well, I am not a salesperson, don't treat anything I say as binding or
C00026 00005 ∂23-Sep-81 2128 Fahlman at CMU-20C Re: Extended PERQ's
C00031 00006 rpg jmc clt lgc jk cg rww rsf jjw moskowski moses ym tob pb
C00032 00007 Mike Genesereth and I have done a head count, finally, of the Lisp community
C00035 00008 ∂02-Oct-81 1410 RPG@Sail (SuNet)
C00038 00009 ∂02-Oct-81 1440 DCL
C00045 00010 ∂04-Oct-81 2152 Daniel L. Weinreb <DLW at MIT-AI> FOO FOO FOO
C00049 ENDMK
C⊗;
The research computing in the Computer Science Department is dominated
by the programming language, Lisp. In particular there are several
ARPA funded research groups who use Lisp nearly exclusively: the Heuristic
Programming Project, Hand Eye, Formal Reasoning, and the Verification project.
In addition, the students of Terry Winograd have all chosen Lisp as the language
for their research.
Every one of these groups have run up against the address space limitations
of the Lisps that they have been using; invariably, these Lisps run on
one of the Digital Equipment Corporation PDP-10 series of computers (KA-10,
KL-10, 2020, or 2060). These machines have 18 bits of address, and since there
are 36 bit words, there are on the order of 2↑18 cons cells available to
the user (256k words). The reality is that this simply isn't enough address
space for the kinds of programs that AI researchers need to experiment with.
Therefore, the Lisp computing needs of the department are at a critical point
at this time. There are several interesting possibilities on the horizon at the
moment in terms of hardware and/or software that will help solve these problems,
and the plan for the department reflects the worth and likely course of events
with respect to these possibilities.
In terms of Lisp dialects, there are two main dialects in contention at the
moment for the attention of the AI community: InterLisp and Common Lisp.
The former is a groundbreaking Lisp in that it introduced the AI community
for the first time to a modern computing environment. With many features to
ease the system building burden, researchers were able to contemplate programs
that had only been dreamt about previously. Much of the existing AI code
at Stanford is written in InterLisp.
Common Lisp is the distillation of the MacLisp and Lisp Machine Lisp
experiences, coupled with the level-headedness of the SpiceLisp design.
MacLisp and its descendants chose the route of efficient implementation to
bring the community a Lisp that could compete with less esoteric
programming languages, such as FORTRAN, in such tasks as severe numeric
computation. However, as time progressed, the lessons of InterLisp
asserted themselves, and a powerful computing environment was adapted.
Many of the surprising language features of Lisp trace their roots to the
innovative (reckless) style of the MacLisp community.
Since there is a string commitment to both branches of the Lisp family tree,
the department is bound to pursue both in the re-equipment plan.
First, there are large, essentially time-shared computers with larger address
spaces that may be able to support Lisp computing. These are: the DEC 2080,
the DEC 11/780, the DEC 11/750, and the LLNL S-1. Of these, the DEC 11/780 and
DEC 11/750 are currently available, but may not have a suitable Lisp running
on them. The DEC 2080 is unavailable (unprototyped) as is the LLNL S-1. These
latter computers may or may not have a suitable Lisp running on them when they
come into existence.
[Brian: Please check these figures since I'm working from a dulled memory
at the moment.]
The DEC 11/780 (Vax) has a 32 bit address space of 8 bit bytes which the hardware
is able to address. This is reduced to 30 bits by operating system needs.
Further, since it takes 2 words per non-cdr-coded cons
cell, there are 27 bits of cons cells available.
A major problem with the the DEC 11/780 is that the page size (512 bytes)
is very small and, consequently, paging overhead, especially in a potentially
non-localized language like Lisp, is enormous. Coupled with the fact that the
machine is a 1 MIPS machine (1/3 of a KL-10), the DEC 11/780 would be expected
to support 1 or 2 Lisp jobs adequately unless many difficult optimizations are
performed. At the current time there is only one Lisp that runs on the Vax,
namely Franz Lisp from Berkeley. The experience with that language is that it
is poorly engineered as a language and that the Vax simply cannot handle more
than 1 such Lisp running at a time. The Masinter report (unpublished) and personal
conversations with Jon L. White at MIT lead one to believe that neither Interlisp
on the Vax nor NIL on the Vax will be much improved over this situation.
The DEC 11/750 is .6 the performance of the DEC 11/780 and so it is not
a serious contender as a time-shared source of Lisp computing. The department
will not consider the purchase of Vaxes as a step in the direction of solving the
Lisp computing problems it faces.
The DEC 2080 is projected to be a very large machine in terms of performance,
being a 24 MIPS machine nominally. Of course, given the small cache memory
it is doubtful that it will achieve more than 10 MIPS in practice on non-numeric
computing. More interestingly, it will support the extended addressing mode
known on the DEC 2060, but with 30 bits of address rather than 23 bits. This
will yield 29 bits of cons cells given a non-cdr-coded scheme. The main
drawback is that there is no Lisp that will suit the needs for the department
that will be able to run on this machine when it exists. The closet contender
is Rutgers Elisp, which is an extended addressing DEC 2060 version of UCILisp.
The mode of addressing is through the mechanism of indirection, and the address
space is broken up into sectors that may be active at any given time. Hedrick
at Rutgers claims that a compiler and an assembler for Lisp could be
constructed that would automate the sectoring of Lisp code.
At this time there is no thought being given to adapting InterLisp to run on this
machine, and there is no thought of Common Lisp achieving this either. However,
Common Lisp is largely written in Common Lisp, and therefore it may be able to
be installed on this machine relatively quickly compared to InterLisp.
Though some software development commitment must be made to this machine
for lisp computing, the department will purchase one DEC 2080
(or an S-1: see below) in 1984 for this purpose.
The LLNL S-1 is a large multi-processor system being constructed at Lawrence
Livermore National Laboratories by, largely, former Stanford people. Much of
the software development is being done by contract to the Stanford Computer
Science Department. There is an excellent chance that the department will
be able to obtain an S-1 uniprocessor within 2 years. The S-1 is a
20 MIPS machine, but with the use of very large cache memories (16k words
of instruction and 16k words of data) there is the likelihood of
obtaining 15 MIPS in practice. The address space is 29 bits of words with
28 bits of cons cells. It is a tagged architecture, which is a very important
feature that Lisp can take advantage of.
Common Lisp (S-1 NIL) will be running on that machine when it becomes available,
and that Lisp will use the specialized numeric instructions, making it a
highly attractive alternative for Robotics and Vision work in Lisp.
However, the machine will run only custom software in that, except for the
Lisp, there will be no attempt to run any code that can run on any machine.
Thus, there are both hardware and software maintenance problems for this
machine.
The decision about whether to purchase an S-1 or a DEC 2080 will be made
in 1983 when there is better information available on them.
For the moment, however, there are relatively bad choices available, and
all of these choices involve the personal computer route. Personal machines
offer advantages in terms of guaranteed performance (once available) and
disadvantages in terms of group interaction in large projects. Nevertheless,
since they are the only sources of Lisp cycles at the moment that satisfy
Lisp address space needs, they must be considered.
At this point there are 2 live alternatives and 1 possibility. The live
alternatives are XEOS (Xerox) Dolphin and the Symbolics Lisp Machine.
Unfortunately, these machines run different Lisps, and the decision about
hardware now must move strongly into the software domain. Dolphins run
InterLisp and Symbolics Lisp Machines run AlphaLisp. InterLisp enjoys a
high software investment at Stanford, centered nearly exclusively on the
HPP project. MacLisp has a high software investment, centered on Formal reasoning
and Hand-eye. The possibility is the Extended PERQ from Three Rivers running
SpiceLisp. Neither SpiceLisp nor the Extended PERQ are available at this time.
There is no point in requiring any re-coding whenever possible, and since
MacLisp is the ancestor of AlphaLisp, Symbolics Lisp Machines suffice
for the non-InterLisp computing. AlphaLisp, moreover, is one of the family
of Common Lisps (S-1 NIL, VAX NIL, and SpiceLisp being the others). So, if
a larger, timeshared machine becomes available for the needs of the non-InterLisp
community which runs a Common Lisp, there will be a small conversion job, if any.
The Dolphins are 1/2 the speed of a KL-10 and have 23 bits of words addressing
capability. The word size is 16 bits, and so 4 words are required per non-cdr-coded
cons cell, meaning that there are 21 bits of cons cells available. However,
there are tag bits available as well, which is an advantage for the Lisp.
The Symbolics ``L'' Lisp Machines are run at 3 times the current Lisp Machines
(CADR) and expect to have the speed of a KL-10, have an address
space of 24 bits, which, due to cdr-coding yields 23.5 bits worth of cons cells.
The L machine comes equipped with 256K 36-bit-plus-ECC words of main memory,
Winchester disk of at least 67 Mbytes, 1100x800 or so B&W screen,
keyboard, mouse, Ethernet II or Chaosnet, 68000 front end processor,
8-bit audio output, and some RS-232 lines. These will run approximately $60k
each.
HPP has estimated that they need 15 Dolphins over the next 3 years. Comparing
group sizes, the department will purchase 8 Symbolics Lisp Machines over the next
two years, at which point, the large timeshared Lisp will become available to
supplement the Lisp Machines. HPP will obtain the 15 Dolphins purchased over
the next 3 years.
[Comment: I've talked to HPP hackers and they report that only the `classic'
HPP stuff (MYCIN...) are still done in InterLisp, and they say they and
their friends in HPP use MacLisp, looking forward to Lisp Machines. What is
going on that they need 15 Dolphins???? We need 30 Lisp Machines, I think.
With Common Lisp on the way and no large timeshared machines for InterLisp
around, aren't we cutting ourselves off here???? Lisp Machines, Vaxes, S-1's
SPICE machines, and maybe the 2080 will run Common Lisp. Am I nuts or are
we being hacked somehow?]
1982 1983 1984 Total
1. S1 $500 $500
2. Dolphins (15) $600 $300 $900
3. Symbolics (8) $240 $240 $480
ITEM 1982 1983 1984 TOTAL
------------------------------------------------------------
1. S1 $500 $500
2. Dolphins (10) 360 240 600
3. Symbolics (5) 325 325
4. Central File Server 225 225
5. SUN work stations(60) 200 50 200 450
6. Low-cost terminals (50) 50 50 100
7. Ethernet equipment 55 55 110
8. Star printers (3) 30 60 90
9. SUVax 11/750 (5) 180 180
10. Local ethernet and local 300 300
file store purchase
and development
11. Maintenance 80 120 200 400
------------------------------------------------------------
TOTAL $1,180 $1,100 $1,000 $3,280
Well, I am not a salesperson, don't treat anything I say as binding or
as having been promised by Symbolics, and please don't tell anybody that
I used the Arpanet for crass, commercial purposes...
By April 1 we will be taking orders for the "Symbolics 3600 Symbol
Processing System", internally known as the L-machine (standing for "Low
cost"). This will be similar to the existing machines, running more or
less the same higher level software (except maybe we will have converted
to Common Lisp by then if the universe smiles on all of us), but with a
new instruction set, new microcode, new compiler, and new debugger, as
well as some new features like an improved GC and upward funargs and
more address space and an attempt at a real-time feature and better
audio output and other stuff. It will be at least three times the basic
CPU speed of the existing Lisp Machine. In the 7 or 8 quantity each
unit cost something on the close order of $60K, consisting of a
processor, 256K 36-bit-plus-ECC words of main memory, Winchester disk of
at least 67 Mbytes, 1100x800 or so B&W screen, keyboard, mouse, Ethernet
II or Chaosnet, 68000 F.E.P., 8-bit audio output, some RS-232 lines and
other assorted little peripherals. This machine does not yet exist, but
SUDS prints are starting to appear and Augat is building us wire-wrap boards
for the prototype; real ones will be PC boards.
For more detailed technical info, please feel free to call us at
617-864-4660. For marketing info, real prices, promises and such, call
the west coast office at 213-347-9224 and ask for either Bob Chisum or
Henry Baker (I am not sure which of them is currently the one for you to
be talking to).
∂23-Sep-81 2128 Fahlman at CMU-20C Re: Extended PERQ's
Date: 24 Sep 1981 0018-EDT
From: Fahlman at CMU-20C
Subject: Re: Extended PERQ's
To: RPG at SU-AI
cc: GLS at SU-AI
In-Reply-To: Your message of 23-Sep-81 2335-EDT
Dick,
The following is not classified, but please don't broadcast the info
without checking with me first -- I might want to hedge and sanitize a
bit more.
There are three Perq extensions under consideration:
1. A 1-Mbyte main memory board to replace the current 256 Kbyte board.
These should be available very soon, for about $8K extra. (All these
prices are 3RCC current plans -- if they have any sense at all they will
lower the somewhat inflated prices quoted here so that the fully
extended machine will come in under $40K.) Since you only get one
memory board in the current cabinet, and since 96K of the current 256K
bytes is for screen map, this extension is essential for any type of
serious work.
2. A new processor board with 16K microstore instead of 4K, and with
corrections to several design errors that were seriously screwing us on
performance in Lisp. We made the decision awhile back that our Lisp
would not attempt to run on the old processor board. The new board will
cost about $8K extra and will be available to customers around the first
of the year. I just got the documentation for the new board today, and
we should be getting a prototype board to play with in three or so
weeks.
3. A hardware page map, probably also with a memory cache. Maybe $6K
extra, but they haven't really started designing it yet, so it's a long
way off. If they do it right, it could buy us a factor of 3 in
non-swap-bound performance.
Given all three of these extensions, the extended Perq will probabaly
qualify as a true, though rather minimal, Spice machine. My guess is
that after some tuning, it wil be maybe a factor of 2-3 slower than the
old MIT Lisp machine. Since both machines will spend most of their time
waiting for page swaps, the real-time differential might be smaller on
big jobs. Then again, you can stuff aditional memory beyond 1 Mbyte
onto a Lisp machine and its disk may be faster, so all such comparisons
get complex. (As you well know.) My guess is that we will have mostly
Perqs here, but a moderate number of Symbolics machines for heavy-duty
AI munching. There should be no real qualitative difference between the
two.
We consider the Dolphin to be in a class with the UNEXTENDED Perq,
except that it allows more main memory; that is, we do not think it is a
good vehicle for Spice, especially for Spice Lisp, and we have no
leverage to get any of the small but necessary changes made. As of
today Dolphin/Interlisp may be the only game in town, but in a year it
will look really pitiful next to Spice/Perq and the Symbolics machine.
-- Scott
-------
rpg jmc clt lgc jk cg rww rsf jjw moskowski moses ym tob pb
dlo dhm klc sl selker jdz pph ml aam dcl wp hjl fwh anderson Malik
rmz
InterLisp Common Lisp
Hpp 24 12
Non-Hpp 4 30
total 28 42
HPP 36
non-Hpp 34
Jan Clayton (M I)
Jeff Finger (M I)
MRG
Milt Grinberg
Nam Quan Huyn
JDM
Chuck Paulson
Steve Tappel
Jay Lark
3 new students (no names)
Mike Genesereth and I have done a head count, finally, of the Lisp community
as it exists at Stanford CSD and CSL. I have included the groups: HPP,
Formal Reasoning, Verification, and Zohar.
The results are a bit surprising: there are 69 Lisp users that we can identify.
Of these there are 41 who use a Lisp other than Interlisp. These will eventually
become Common Lisp users. 28 are InterLisp users. Of the entire Lisp community,
36 are HPP and 33 are non-HPP. Within HPP 24 use InterLisp and 12 non-InterLisp.
InterLisp Common Lisp
Hpp 24 12
Non-Hpp 4 29
Total 28 41
HPP 36
Non-HPP 33
I think that the responsible thing for us to do is to equip the Lisp community
according to this ratio; if individual groups want to supply otherwise,
then they may.
This means that of the 15 lisp machines, we should buy 9 Symbolics machines
and 6 Dolphins. 3 Symbolics machines will be for HPP primary use and
6 for non-HPP primary use. All the Dolphins will be for HPP primary use.
I know that this wmight not sit well with everyone on the committee,
but I am willling to listen to arguments against equity.
About future Lisp computing needs. When a large processor capable of running Lisp
well (e.q. an S-1) arrives, we should consider re-distributing things, though
I imagine that there will continue to be a need for Lisp cycles of all types.
-rpg-
∂02-Oct-81 1410 RPG@Sail (SuNet)
Date: 02 Oct 1981 1404-PDT
From: Dick Gabriel <RPG at SU-AI>
To: equip at DIABLO, (sunet) at Sail, csd.genesereth at SU-SCORE
Mike Genesereth and I have done a head count, finally, of the Lisp community
as it exists at Stanford CSD and CSL. I have included the groups: HPP,
Formal Reasoning, Verification, and Zohar.
The results are a bit surprising: there are 60 Lisp users that we can identify.
Of these there are 36 who use a Lisp other than Interlisp. These will eventually
become Common Lisp users. 24 are InterLisp users. Of the entire Lisp community,
36 are HPP and 24 are non-HPP. Within HPP 24 use InterLisp and 12 non-InterLisp.
InterLisp Common Lisp
Hpp 24 12
Non-Hpp 0 24
Total 24 36
HPP 36
Non-HPP 24
I think that the responsible thing for us to do is to equip the Lisp community
according to this ratio; if individual groups want to supply otherwise,
then they may.
This means that of the 15 lisp machines, we should buy 9 Symbolics machines
and 6 Dolphins. 3 Symbolics machines will be for HPP primary use and
6 for non-HPP primary use. All the Dolphins will be for HPP primary use.
I know that this wmight not sit well with everyone on the committee,
but I am willling to listen to arguments against equity.
About future Lisp computing needs. When a large processor capable of running Lisp
well (e.q. an S-1) arrives, we should consider re-distributing things, though
I imagine that there will continue to be a need for Lisp cycles of all types.
-rpg-
∂02-Oct-81 1440 DCL
∂02-Oct-81 1250 RPG Lisp
Does your group still use Lisp? Do you plan to use Lisp in your research?
If so, how many people are Lisp users in your group? If you intend to
use a Vax of some sort as a Lisp machine, you ought to talk to me first,
since experience has it that Vaxes are atrocious Lisp machines.
-rpg-
REPLY:
YES ABSOLUTELY.
We are committed to using LISp as our main implementation language for at
least another two years.
- David
Then, we ought to chat about your exact needs wrt Lisp. I am assuming you
have 4 full time Lisp users.
∂02-Oct-81 1529 DCL Hardware wish list
To: RPG at SU-AI
CC: REG at SU-AI, EAF at SU-AI
CC: Reid at SU-SCORE, csd.cheriton at SU-SCORE,
csl.jlh at SU-SCORE, DCL at SU-AI
Regarding the ARPA wish list for hardware:
My thinking on this is that industry currently has under wraps
some new machine/software systems that will be available in 2-3 yrs.
and will make vax 750 workstations look silly, both performance wise
and cost wise. I place both IBM and HP in this class.
Given this, my immediate concern is that we get systems that enable
developed and current software work on my projects to continue without
disruption this year and next year.
Therefore I would vote for a large KL20-60 system for CSL.
Various SUN workstations and other graphics machines should still be a
a good idea from my point of view.
A lisp machine this year would be ok too.
-David
∂02-Oct-81 2110 RPG@Sail (SuNet) David Luckham's wishlist
Date: 02 Oct 1981 2108-PDT
From: Dick Gabriel <RPG at SU-AI>
Subject: David Luckham's wishlist
To: equip at DIABLO, (sunet) at Sail
I asked David what he needed to continue his work wrt Lisp after
I saw the CSL wishlist. I found it hard to believe that he wanted
Vaxes etc if he were still using Lisp heavily. Cheriton seemed to
think David was cognizant of things and that Vaxes was what he wanted.
A 2060 allows the Verification group to use MacLisp in a slightly
faster, more flexible environment than on SAIL. He also mentioned a
Lisp machine. In my accounting of Lisp users his group was counted,
and I assume he would use a Lisp Machine quite nicely for his work.
In my previous message about Lisp I talked about a 3/2 ratio of Symbolics
machines to Dolphins as the target for the ARPA re-equipment proposal.
I believe, though, that HPP might need further Dolphin enhancement,
unless there is some motion towards Common Lisp from InterLisp there
in the future. I think that given budget constraints HPP should acquire
these additional Dolphins with other money since the equipment committee
must address a larger community of Lisp users. If I had my real choice
I would have asked for 15 Dolphins and 24 Symbolics Machines, totalling
$3.1mbucks.
Someone mentioned eliminating the S-1 from the budget. This certainly, then,
has to mean that Lisp Machines replace the S-1, and in that case I would
have to ask for 39 Lisp machines and Dolphins over 3 years.
∂02-Oct-81 2331 Wiederhold at SUMEX-AIM Re: Lisp use
Date: 2 Oct 1981 2327-PDT
From: Wiederhold at SUMEX-AIM
Subject: Re: Lisp use
To: RPG at SU-AI
cc: kaplan at SRI-KL
In response to your message sent 02 Oct 1981 1251-PDT
We have been using interlisp, rather because of its ubiquitessness
here than for intrinsic reasons.
I believe that over time we can adapt to any LISP, and a common lisp
seems desirable. Our projects are prob. a bit less hacky than those of
the HPP bunch.
We do need a good file interface, have bben working on one for Interlisp-
we are tlaking here about real database style files.
I cannot judge the difficulty of marryingbthis work to any other other LISP.
the file package is written in PASCAL, and the problems are in
scheduling th forks to cooperate and share data.
Perhaps jerry can add comments.
I will value highly the possibility of doing LISP and other languages on one machine,
and for that reason am not taken with Dolphins.
Gio
-------
∂04-Oct-81 2152 Daniel L. Weinreb <DLW at MIT-AI> FOO FOO FOO
Date: 5 October 1981 00:49-EDT
From: Daniel L. Weinreb <DLW at MIT-AI>
Subject: FOO FOO FOO
To: RPG at SU-AI
We understand that the capability is important. I wish we were such a
position that we could promise the capabilty. Unfortunately, we are not --
we would be bullshitting you if we said "yeah, we'll be able to do that".
In theory, you could redo the window system that way, in any of a few
different ways. The "right" way would be to make a new flavor that
took all the window system messages and sent messages using a protocol
rather than affecting the bit-map. Unfortunately, lots of the system does not
go through the message-passing interface, for efficiency reasons; some
things even call directly into microcode. On the L-machine we will
evntually make message passing much more efficient, but it would still
be a lot of work to go back and fix everything, and I don't know how long it
would take for us to get around to that since there will be more important
things to do. Anything without which L machines simply won't work at all
must have higher priority than things that add new functionality, like ability
to work in the environment you outline. It's not impossible to win on
the L machine, but we just can't commit the resources at this point.
"What can we do to convince you..." Well, if you provided us with another
super-hacker who could be assigned to work on this project, someone good
enough that HIC and MMcM would be willing to let him at the guts of the
window system, and work out the logistics of getting him to Boston and
paying for him...
Well, this message is nicer than the last in that it gives a hint that it
is eventually possible rather than inherently impossible. Consider
that when we eventually get the machines, the networks etc, we may very
well be able to have me, for example, work on a version of this for our
use. Your answer, thus, that the hardware doesn't prevent it cheers me up.
In any event, I am hoping to be able to order 5-10 L machines in
the relatively near future, if all goes well with my political battles
with Feigenbaum.
-rpg-