> From: "Theodore Y. Ts'o"
> To be fair, it's really easy to be wise to after the fact.
Right, which is why I added the caveat "seen to be such _at the time_ by some
people - who were not listened to".
> failed protocols and designs that collapsed of their own weight because
> architects added too much "maybe it will be useful in the future"
And there are also designs which failed because their designers were too
un-ambitious! Converting to a new system has a cost, and if the _benefits_
(which more or less has to mean new capabilities) of the new thing don't
outweigh the costs of conversion, it too will be a failure.
> Sometimes having architects being successful to add their "vision" to a
> product can be worst thing that ever happened
A successful architect has to pay _very_ close attention to both the 'here and
now' (it has to be viable for contemporary use, on contemporary hardware, with
contemporary resources), and also the future (it has to have 'room to grow').
It's a fine edge to balance on - but for an architecture to be a great
success, it _has_ to be done.
> The problem is it's hard to figure out in advance which is poor vision
> versus brilliant engineering to cut down the design so that it is "as
> simple as possible", but nevertheless, "as complex as necessary".
Absolutely. But it can be done. Let's look (as an example) at that IPv3->IPv4
addressing decision.
One of two things was going to be true of the 'Internet' (that name didn't
exist then, but it's a convenient tag): i) It was going to be a failure (in
which case, it probably didn't matter what was done, or ii) it was going to be
a success, in which case that 32-bit field was clearly going to be a crippling
problem.
With that in hand, there was no excuse for that decision.
I understand why they ripped out variable-length addresses (I was just about
to start writing router code, and I know how hard it would have been), but in
light of the analysis immediately above, there was no excuse for looking
_only_ at the here and now, and not _also_ looking to the future.
Noel
> From: Tony Finch <dot(a)dotat.at>
> Was this written down anywhere?
Alas, no. It was a presentation at a group seminar, and used either hand-drawn
transparencies, or a white-board - don't recall exactly which. I later tried to
dig it up for use in Nimrod, but without success.
As best I now recall, the concept was that instead of the namespace having a
root at the top, from which you had to allocate downward (and then recurse),
it built _upward_ - if two previously un-connected chunks of graph wanted to
unite in a single system, they allocated a new naming layer on top, in which
each existing system appeared as a constituent.
Or something like that! :-)
The issue with 'top-down' is that you have to have some global 'authority' to
manage the top level - hand out chunks, etc, etc. (For a spectacular example
of where this can go, look at NSAP's.) And what do you do when you run out of
top-level space? (Although in the NSAP case, they had such a complex top
couple of layers, they probably would have avoided that issue. Instead, they
had the problem that their name-space was spectacularly ill-suited to path
selection [routing], since in very large networks, interface names
[adddresses] must have a topological aspect if the path selection is to
scale. Although looking at the Internet nowadays, perhaps not!)
'Bottom-up' is not without problems of course (e.g. what if you want to add
another layer, e.g. to support potentially-nested virtual machines).
I'm not sure how well Dave understood the issue of path selection scaling at
the time he proposed it - it was very early on, '78 or so - since we didn't
understand path selection then as well as we do now. IIRC, I think he was
mostly was interested in it as a way to avoid having to have an asssignment
authority. The attraction for me was that it was easier to ensure that the
names had the needed topological aspect.
Noel
> From: Dave Horsfall
> idiots keep repeating those "quotes" ... in some sort of an effort to
> make the so-called "experts" look silly; a form of reverse
> jealousy/snobbery or something? It really pisses me off
You've just managed to hit one of my hot buttons.
I can't speak to the motivations of everyone who repeats these stories, but my
professional career has been littered with examples of poor vision from
technical colleagues (some of whom should have known better), against which I
(in my role as an architect, which is necessarily somewhere where long-range
thinking is - or should be - a requirement) have struggled again and again -
sometimes successfully, more often, not.
So I chose those two only because they are well-known examples - but, as you
correctly point out, they are poor examples, for a variety of reasons. But
they perfectly illustrate something I am _all_ too familiar with, and which
happens _a lot_. And the original situation I was describing (the MIT core
patent) really happened - see "Memories that Shaped an Industry", page 211.
Examples of poor vision are legion - and more importantly, often/usually seen
to be such _at the time_ by some people - who were not listened to.
Let's start with the UNIBUS. Why does it have only 18 address lines? (I have
this vague memory of a quote from Gordon Bell admitting that was a mistake,
but I don't recall exactly where I saw it.) That very quickly became a major
limitation. I'm not sure why they did it (the number of contact on a standard
DEC connector probably had something do with it, connected to the fact that
the first PDP-11 had only 16-bit addressing anyway), but it should have been
obvious that it was not enough.
And a major one from the very start of my career: the decision to remove the
variable-length addresses from IPv3 and substitute the 32-bit addresses of
IPv4. Alas, I was too junior to be in the room that day (I was just down the
hall), but I've wished forever since that I was there to propose an alternate
path (the details of which I will skip) - that would have saved us all
countless billions (no, I am not exaggerating) spent on IPv6. Dave Reed was
pretty disgusted with that at the time, and history shows he was right.
Dave also tried to get a better checksum algorithm into TCP/UDP (I helped him
write the PDP-11 code to prove that it was plausible) but again, it got turned
down. (As did his wonderful idea for bottom-up address allocation, instead of
top-down. Oh well.) People have since discussed issues with the TCP/IP
checksum, but it's too late now to change it.
One place where I _did_ manage to win was in adding subnetting support to
hosts (in the Host Requirements WG); it was done the way I wanted, with the
result that when CIDR came along, even though it hadn't been forseen at the
time we did subnetting, it required _no_ hosts changes of any kind. But
mostly I lost. :-(
So, is poor vision common? All too common.
Noel
On 2018-06-18 04:00, Ronald Natalie <ron(a)ronnatalie.com> wrote:
>> Well, it is an easy observable fact that before the PDP-11/70, all PDP-11 models had their memory on the Unibus. So it was not only "an ability at the lower end", b
> That’s not quite true. While the 18 bit addressed machines all had their memory directly accessible from the Unibus, the 45/50/55 had a separate bus for the (then new) semiconductor (bipolar or MOS) memory.
Eh... Yes... But...
The "separate" bus for the semiconductor memory is just a second Unibus,
so the statement is still true. All (earlier) PDP-11 models had their
memory on the Unibus. Including the 11/45,50,55.
It's just that those models have two Unibuses.
Johnny
--
Johnny Billquist || "I'm on a bus
|| on a psychedelic trip
email: bqt(a)softjar.se || Reading murder books
pdp is alive! || tryin' to stay hip" - B. Idol
We lost the Father of Computing, Alan Turing, on this day when he suicided
in 1954 (long story). Just imagine where computing would've been now...
Yes, there are various theories surrounding his death, such as a jealous
lover, the FBI knowing that he knew about Verona and could be compromised
as a result of his sexuality, etc. Unless they speak up (and they ain't),
we will never know.
Unix reference? Oh, that... No computable devices (read his paper), no
computers... And after dealing with a shitload of OSs in my career, I
daresay that there is no more computable OS than Unix. Sorry, penguins,
but you seem to require a fancy graphical interface these days, just like
Windoze.
--
Dave Horsfall DTM (VK2KFU) "Those who don't understand security will suffer."
> From: Derek Fawcus <dfawcus+lists-tuhs(a)employees.org>
> my scan of it suggests that only the host part of the address which were
> extensible
Well, the division into 'net' and 'rest' does not appear to have been a hard
one at that point, as it later became.
> The other thing obviously missing in the IEN 28 version is the TTL
Yes... interesting!
> it has the DF flag in the TOS field, and an OP bit in the flags field
Yeah, small stuff like that got added/moved/removed around a lot.
> the CIDR vs A/B/C stuff didn't really change the rest.
It made packet processing in routers quite different; routing lookups, the
routing table, etc became much more complex (I remember that change)! Also in
hosts, which had not yet had their understanding of fields in the addresses
lobotomized away (RFC-1122, Section 3.3.1).
Yes, the impact on code _elsewhere_ in the stack was minimal, because the
overall packet format didn't change, and addresses were still 32 bits, but...
> The other bit I find amusing are the various movements of the port
> numbers
Yeah, there was a lot of discussion about whether they were properly part of
the internetwork layer, or the transport. I'm not sure there's really a 'right'
answer; PUP:
http://gunkies.org/wiki/PARC_Universal_Packet
made them part of the internetwork header, and seemed to do OK.
I think we eventually decided that we didn't want to mandate a particular port
name size across all transports, and moved it out. This had the down-side that
there are some times when you _do_ want to have the port available to an
IP-only device, which is why ICMP messages return the first N bytes of the
data _after_ the IP header (since it's not clear where the port field[s] will
be).
But I found, working with PUP, there were some times when the defined ports
didn't always make sense with some protocols (although PUP didn't really have
a 'protocol' field per se); the interaction of 'PUP type' and 'socket' could
sometimes be confusing/problemtic. So I personally felt that was at least as
good a reason to move them out. 'Ports' make no sense for routing protocols,
etc.
Overall, I think in the end, TCP/IP got that all right - the semantics of the
'protocol' field are clear and simple, and ports in the transport layer have
worked well; I can't think of any places (other than routers which want to
play games with connections) where not having ports in the internetwork layer
has been an issue.
Noel
> From: Derek Fawcus
> Are you able to point to any document which still describes that
> variable length scheme? I see that IEN 28 defines a variable length
> scheme (using version 2)
That's the one; Version 2 of IP, but it was for Version 3 of TCP (described
here: IEN-21, Cerf, "TCP 3 Specification", Jan-78 ).
> and that IEN 41 defines a different variable length scheme, but is
> proposing to use version 4.
Right, that's a draft only (no code ever written for it), from just before the
meeting that substituted 32-bit addresses.
> (IEN 44 looks a lot like the current IPv4).
Because it _is_ the current IPv4 (well, modulo the class A/B/C addressing
stuff). :-)
Noel
> From: Johnny Billquist
> It's a separate image (/netnix) that gets loaded at boot time, but it's
> run in the context of the kernel.
ISTR reading that it runs in Supervisor mode (no doubt so it could use the
Supervisor mode virtual address space, and not have to go crazy with overlays
in the Kernet space).
Never looked at the code, though.
Noel
On 2018-06-17 04:00, jnc(a)mercury.lcs.mit.edu (Noel Chiappa) wrote:
> > From: Johnny Billquist
>
> > It's a separate image (/netnix) that gets loaded at boot time, but it's
> > run in the context of the kernel.
>
> ISTR reading that it runs in Supervisor mode (no doubt so it could use the
> Supervisor mode virtual address space, and not have to go crazy with overlays
> in the Kernet space).
Yes. That rings a bell now that you mention it. Pretty sure you are correct.
Johnny
--
Johnny Billquist || "I'm on a bus
|| on a psychedelic trip
email: bqt(a)softjar.se || Reading murder books
pdp is alive! || tryin' to stay hip" - B. Idol
> From: Johnny Billquist
> incidentally have 18 data bits, but that is mostly ignored by all
> systems. I believe the KS-10 made use of that, though. And maybe the
> PDP-15.
The 18-bit data thing is a total kludge; they recycled the two bus parity
lines as data lines.
The first device that I know of that used it is the RK11-E:
http://gunkies.org/wiki/RK11_disk_controller#RK11-E
which is the same cards as the RK11-D, with a jumper set for 18-bit operation,
and a different clock crystal. The other UNIBUS interface that could do this
was the RH11 MASSBUS controller. Both were originally done for the PDP-15;
they were used with the UC15 Unichannel.
The KS10:
http://gunkies.org/wiki/KS10
wound up using the 18-bit RH11 hack, but that was many years later.
Noel
