Google Groups no longer supports new Usenet posts or subscriptions. Historical content remains viewable.
Dismiss

Re: Named Memory

293 views
Skip to first unread message

Louis Krupp

unread,
Jun 15, 2014, 5:30:49 AM6/15/14
to
On Fri, 13 Jun 2014 09:41:29 -0700, Ivan Godard <iv...@ootbcomp.com>
wrote:

>On 6/13/2014 2:56 AM, Louis Krupp wrote:
>> On Thu, 12 Jun 2014 21:54:52 -0700 (PDT), Seima Rao
>> <seim...@gmail.com> wrote:
>>
>>> On Wednesday, June 11, 2014 7:43:01 PM UTC+5:30, Piotr Wyderski wrote:
>
>> The Burroughs Large Systems operating system used to optimize access
>> to elements of structures by setting a display register to the address
>> of the structure and then using address couples (lex level, offset)
>> to treat structure components as regular local variables (what you're
>> calling named objects). The current Unisys successor probably does
>> the same thing.
>
>Thank you - I hadn't known about doing that trick, it must have been
>introduced after my day. They must have added an instruction to set the
>display - do you have details?

I did. Once.

I remember the trick being used for the File Information Block, which
when could be accessed as an array or as a stack with local variables.
A procedure called FIBSTACK was involved.

A few years ago, someone kindly sent me a text file with the source
for the 2.1 (ca 1970?) MCP. I've found FIBSTACK, and while some of
the code looks familiar, a lot of it is now beyond comprehension. What
does PCWSTORE do? Is it an ESPOL intrinsic that I can't find in any
period ESPOL manuals? I've searched for its definition in the MCP
source, but I can't find it there, either.

I'm cross-posting this to comp.sys.unisys; there are folks there who
know this stuff better than I. I've also set follow-ups to
comp.sys.unisys, as this has as much to do with software as with
hardware, and it's unlikely that it will be of too much interest to
most of the readers of comp.arch.

I remember reading somewhere that modern ALGOL Structure Blocks might
also be accessible as a stack with local variables. Can anyone on
comp.sys.unisys clarify?

Louis

Paul Kimpel

unread,
Jun 16, 2014, 9:01:01 AM6/16/14
to
On the contrary, readers of comp.arch should be vitally interested about
FIBSTACK, since it's all about architecture.

The Burroughs B6x00/B7x00 systems and their successors to the present
have always had instructions to directly read (RPRR) and write (SPRR)
the D (display) registers, but these were rarely used outside of
diagnostic software, and they were not necessary to implement the type
of addressing that Louis describes. The registers are set by the normal
display register update that is done during a procedure call.

There are three ideas behind this concept:

1. A "stack" on the Burroughs systems is three things: (a) a push-pop
workspace for expression evaluation, (b) a storage area for
procedure-call history, and (c) and addressing environment for the
nested scopes needed by a block-structured programming language. That
last is the most important to this discussion.

2. A stack frame that implements the storage for locals of a procedure
is just a contiguous chunk of words in memory. If the stack frame is
part of the current execution scope, then there will be a D register
pointing to the first word of that frame. Therefore, the words of that
frame can be addressed as locals using normal address couples,
(LL,offset), where LL (lexicographic level) is the D register index and
offset is the displacement from the base of the stack frame to the word
desired. This is the most efficient form of addressing available to
Normal State programs.

3. In a block-structured language such as Algol or Pascal, there are D
registers pointing not only to the base of the current stack frame, but
to the bases of all of the lower-level stack frames that are currently
in scope. All of their words are also accessible using that same
efficient address-couple mechanism. Therefore, all you have to do is
arrange for the data structure you want to access to look like a
more-global stack frame of a procedure you call.

Also note that procedures can be called by indirect references (e.g., a
procedure identifier passed by name) and that a more-global stack frame
does not need to be part of the same program's stack. The Burroughs
architecture implemented cross-stack addressing (sometimes called
"cactus stack" addressing). See Hauck and Dent's 1968 AFIPS paper for
details (e.g.,
http://www.cs.berkeley.edu/~culler/courses/cs252-s05/papers/burroughs.pdf).
Also, the B6700 Reference Manual is available at
http://bitsavers.org/pdf/burroughs/B6500_6700/ (in particular, see
Section 3).

The original idea behind this type of addressing probably was to
implement hierarchical process families, but it turned out to have a
number of other useful applications as well. It just took a while to
appreciate them.

In the general case, the setup and activation of this type of addressing
can get really complex, and even the simple cases are not all that
simple. There are three control words that do most of the heavy lifting,
The Stuffed Indirect Reference Word (SIRW), the Mark Stack Control Word
(MSCW) and the Program Control Word (PCW).

I'll describe how these worked on the B6x00/B7x00 systems. The details
have changed quite a bit over the years, but the concept has remained
the same. In the following, the notation [S:N] defines a bit field
within a word, with S denoting the starting bit number (high order=47,
low-order=0) and L denoting the number of bits to the right (towards bit
0) from the starting bit.

An SIRW (tag=1) is used to point to a word in a stack. Its main use is
to point to words that are (or may be) outside the current scope chain.
[There used to be an un-stuffed IRW that pointed only to words in the
current scope; it is no longer present in modern systems]. It consists
of three main fields.

[45:10] the stack number of the stack in which the word exists.
Each stack is defined by a data descriptor having its base address and
length. All of these stack descriptors are arranged in a one-dimensional
array, the Stack Vector, indexed by their stack number. A descriptor for
the Stack Vector is stored at (0,2) in the MCP stack. This is an address
couple known to the hardware.

[35:16] the displacement from the base of the stack identified
above to the start of the stack frame containing the word being
addressed. The word at the base of that stack frame must be an MSCW.

[12:13] the offset from the start of that stack frame to the word
being addressed. Thus, if you add this offset and the displacement
above, that yields the offset to the word from the base of stack. The
two values are stored separately because there are cases (particularly
during procedure call) where you need to determine the scope within
which the referenced word exists.

An MSCW (tag=3) provides stack frame and scope linkage. Its major fields
are:

[45:10] stack number of the stack that contains the stack frame for
the next lower-level (more-global) scope.

[35:16] displacement from the base of that stack to the start of
the stack frame for the next more-global scope. The word at the start of
each stack frame must be an MSCW. Thus, this field and the stack number
link to the stack frames for the more-global scopes, and allow the
processor to restore the D registers for the scope chain.

[18:5] lexicographic level associated with this stack frame.

[12:13] offset to the prior stack frame in this stack. This links
to the stack frame of the immediate caller, and allows the stack to be
cut back when a procedure exits.

A PCW (tag=7) defines the entry point to a procedure or subroutine. Its
major fields are:

[45:10] stack number of the stack containing this PCW.

[35:3] the byte offset within the first word of code where
execution is to begin.

[32:13] the word offset within the code segment for the procedure
where execution is to begin.

[18:5] the lexicographic level at which the procedure will run.

[12:13] the displacement within the program's Segment Dictionary to
the descriptor for the procedure's code segment. The Segment Dictionary
is itself a stack, with its own stack number, but is used only as a
read-only addressing environment. It consists of a single stack frame,
which for Normal State programs is always addressed at LL=1 (the global
data environment for most use programs is LL=2, and the MCP globals are
at LL=0). The Segment Dictionary contains descriptors for code segments,
descriptors for read-only data segments (e.g., Algol Value Arrays), and
single-word constants. Since the Segment Dictionary is at a more global
level, contains only read-only elements, and is implemented as a
separate stack, it is sharable among multiple data stacks. Hence,
universal and automatic code reentrancy.

Thus, the PCW indicates the segment, word, syllable (byte), and lex
level of the entry point to the procedure.

The more subtle thing about a PCW is that its _location in the stack_
determines the addressing scope for the procedure. The stack frame in
which the PCW resides becomes the next more-global scope for the called
procedure. The MSCW for that stack frame then links to the scopes that
are more global than that, if any.

This is the key to getting the D registers to address data areas that
would not normally be considered to be part of a stack. If we can put an
MSCW at the beginning of the area that contains appropriate linkage
values, then place PCWs within that area, then construct SIRWs that
point to those PCWs, and use those SIRWs in a procedure-call sequence,
the called procedure will have the data area as its next more-global
addressing environment, and can access the words of that area using
efficient address-couple addressing. What's more, the MSCW at the
beginning of that data area can be linked to yet more-global
environments, so you could implement a nicely hierarchical addressing
scheme for multiple data areas.

To me, the really cool thing is that all of the register setup and
restoration is completely automatic. Procedure call automatically loads
the D registers with the necessary values. Procedure exit automatically
restores the original state. Don't need no stinkin' instructions to set
registers.

A typical procedure call sequence would look like this:

MKST mark stack: construct and push a MSCW
NAMC (LL,n) name call: push an IRW to the PCW or its SIRW
(code to push any parameters to the procedure)
ENTR enter the procedure

That's it -- four bytes of code, plus whatever is necessary to evaluate
the parameters.

The FIBSTACK mechanism that Louis mentioned was the first (that I know
of) to exploit this addressing technique, and it's a good example to
explore. It was implemented as part of the complete rewrite of Logical
I/O (LIO) that was released in the Mark II.0 MCP. LIO is the user-level
API used to access files. The original implementation was based on that
of the B5500, but was an ambitious generalization of many of the B5500
I/O concepts, including the idea of File Attributes. Alas, the original
implementation was complex, buggy, and really, really slow.

Part of the problem was that with all of that generalization, there were
a lot of features and options to choose from, and if you tried to
implement selection of the necessary features in a straightforward way,
it took a lot of tests to figure out what a particular I/O call had to
do. The solution was to break up that one big, mother READ routine into
a bunch of small, tightly focused routines that each addressed a
specific situation, and to build a mechanism that could efficiently
select and execute the appropriate one in any given case.

A FIB (File Information Block) is a system-managed data structure that
represents a file declared in a program. Note that we generally use
"file" to represent two entirely different, but related things. There is
"file" meaning a physical collection of data -- a reel of tape, a set of
areas or extents on disk, etc. Then there is "file" meaning the API we
use to access those physical entities. This discussion is oriented
towards that latter meaning.

Physically, a FIB is just an array of words in memory. A file
declaration in a program is represented at run time by a data descriptor
for that array. Like most such things in the B5500 and successor
systems, it is allocated on first reference. Another data structure
stored in the code file and addressed through the Segment Dictionary
contains a list of attribute/value pairs used to initialize the FIB and
set values established by the file declaration. As part of that
initialization, any attributes specified by file equations in the
initiating environment (e.g., FILE statements in WFL) will be merged
with the compiled-in attributes. User programs do not manipulate the
contents of the FIB directly; they do that through file attribute
assignments and I/O statements.

The LIO routines in the MCP do access the FIB array directly, however,
and in the original implementation that was done by indexing the FIB
descriptor, a more expensive form of addressing than that of address
couples. Thus, a second goal in the redesign was to give the LIO
routines much more efficient access to the words in the FIB.

The original way this was done on the B6500/6700 was to create an
additional set of descriptors in the Stack Vector that effectively
mapped all of physical memory. A stack for that version of the
architecture could be up to 64Ki words in size, and maximum memory was
1Mi words, so only 16 additional descriptors were needed.

The MCP initialized the FIB array so that FIB[1] contained an MSCW. This
was a dummy MSCW, in that there was no prior procedure call, and no
more-global scope. It was there to satisfy the addressing requirements
of the hardware so that the D registers could be loaded and restored
properly.

FIB[0] contained an SIRW termed the "Selector". This SIRW had a stack
number corresponding to the 64KiW chunk of memory in which the FIB was
allocated, and a displacement in [35:16] from the beginning of that
chunk to the MSCW in FIB[1]. The offset field in [12:13] pointed into
the FIB, but the low-order three bits were zero. This SIRW pointed to a
PCW for a sequential WRITE routine. The next word was a PCW for a
sequential READ routine; the word after that was for a random WRITE
routine, etc. There were eight such procedures; modern systems support
up to 16.

To perform an I/O call, the code in the user program generated by the
compiler would first mark the stack. Then it would index and load the
word at FIB[0] onto the top of stack. It would then OR into the
low-order bits of the SIRW the offset for the particular type of I/O
call being done (0=sequential write, 1=sequential read, 2=random write,
3=random read, etc.) It could not do an ADD to apply this offset, since
the SIRW has tag=1 and is not considered to be a data operand, but the
LOR operator would tolerate the non-data tag.

Now we have an SIRW on top of stack, above the MSCW pushed by the
mark-stack operator, which points to a PCW in the FIB. The
compiler-generated code in the user program completes the LIO call by
pushing parameters into the stack (a value used to specify random record
address/carriage control, a length, an indexed descriptor to the
program's record area, etc.) An ENTR operator completes the call by
pushing a return address onto the stack, using the linkage in the SIRW
to traverse the new scope chain and load the D registers as necessary,
load the new code address from the PCW selected in the FIB, and start
executing the code for the LIO procedure.

Note that the FIB itself _is not passed as a parameter_ in the call.
Instead, the fact that the PCW was located in the pseudo-stack frame
contained within the FIB makes the area of memory starting with the
FIB's MSCW the immediately global scope for that procedure, and
therefore the LIO procedure can access words of the FIB using address
couples having offsets relative to the MSCW. In Algol terms, the FIB
looks to the procedure like that procedure's containing block.

Another aspect of this implementation, and one that probably had more
impact on performance than the clever addressing scheme, was that the
set of PCWs indexed in the FIB were mutable. There were many dozens of
LIO procedures, possibly hundreds. Each procedure that was optimized for
a specific I/O situation, and contained as little internal testing and
branching as possible. There were, for example, separate routines for
blocked and unblocked I/O, translated vs. untranslated I/O, tape vs.
disk vs. printer I/O. If the conditions under which I/O was being
performed changed (say, switching between sequential and random reads),
even while the file was open, LIO would replace certain PCWs with others
that were more suitable for the new conditions. This behind-the-scenes
adaptation was transparent to user programs.

Even though the refactoring of LIO into lots of smaller, streamlined
procedures was responsible for most of the performance gain, it was the
architectural support in the system that allowed that approach to be
implemented so efficiently.

As an aside, the master set of LIO PCWs was stored in an array within
the MCP. I became intimately familiar with that array during one
episode, probably in late 1971. We had just installed the Mark II.0 MCP,
which not only included the first release of FIBSTACK Logical I/O, but
an equally re-engineered physical I/O implementation. Our B6500 started
throwing fatal dumps at random intervals. None of us knew anything about
FIBSTACK at that point, just that we had a new (and MUCH faster) I/O
implementation. Analysis of the dumps (at that time still done on an
8.5x11x11 stack of paper with a red pen) showed the processor trying to
do a procedure call against a word that had a tag of 5 (data
descriptor), not 7 (PCW), so the question was, why are we trying to do a
procedure call against a data descriptor? The dumps also showed that it
was supposed to be a PCW in a FIB. After a lot of time pouring over the
MCP listing and tracing things back in the dumps, we found the master
array of LIO PCWs, and lo, *ALL* of them had tags of 5. So the next
question was, what could cause all of the words in an array to suddenly
change from tag 7 to tag 5? The answer, some days later, turned out to
be overlay (virtual memory) I/O. That master array of PCWs was
overlayable, and memory allocation pressure was occasionally forcing it
out to disk. We eventually discovered that the problem only occurred
when the overlay I/O was done by the fourth of the four Multiplexor I/O
channels on the system. That channel would be selected only if the other
three were already busy, which up to that point they probably never had
been. The new LIO and physical I/O implementations were so much more
efficient, though, that the system started doing many more I/Os in
parallel, and channel #4 started to be used occasionally. Once we
determined the problem was specific to that channel, we turned the
problem over to the FEs, who eventually found a wire missing from the
backplane. And yes, that wire carried a signal for the second bit of the
tag. One quick wirewrap job later, the dumps disappeared.

40+ years later, the FIBSTACK mechanism is still present in the modern
Unisys ClearPath MCP, although a lot of low-level hardware details have
changed, largely to accommodate larger stack numbers and memory address
widths. It has held up well against the significant changes in I/O
technology that have occurred over the years. As best I can tell from a
fading memory and samples of current compiler code generation, the
calling sequence in user programs has not changed at all. For example,
here is modern code for an Algol sequential WRITE statement:

WRITE (F, 100, A);
MKST AE mark the stack
ZERO B0 push zero index for the FIB
NAMC (02,0004) 5004 push SIRW to the FIB descriptor
STFF AF stuff the SIRW (these days it's a no-op)
INDX A6 index the FIB descriptor by the zero
LOAD BD push the word at FIB[0] onto the stack
ZERO B0 push record/CC (not used with sequential)
LT8 100 B264 push the number of units to write
ZERO B0 push zero index for the record area
NAMC (02,0003) 5003 push SIRW to the user's record area
INDX A6 index to element zero of record area
LT48 BE 080000000080 (push some control data)
ENTR AB enter the LIO procedure

Note that since this is a sequential write at offset zero within the
FIB's PCWs, no offset had to be OR-ed into the Selector SIRW at FIB[0].
For other operations, that code would be between the first INDX operator
and LOAD. For example, if this were a random read (which uses offset 3),
these additional instructions would appear:

LT8 3 B203 push a literal 3
LOR 91

Algol Structure Blocks are indeed similar to the FIBSTACK mechanism, and
are essentially a generalization of it that is safe to use in user-level
code. They are something of an object-oriented mechanism. You define a
Structure Block as a "type" (somewhat like an OO class), then use that
type to declare either instances or arrays of instances for that type.
Allocation of memory for an instance occurs on first reference. There is
no "new" operator or its equivalent.

Structure Blocks have a body, which can contain data and procedure
declarations, similar to the declarative part of a standard Algol block.
Each declaration can be marked as public (visible outside the block) or
private. Most Algol declarations are allowed within a Structure Block,
but there are some restrictions, e.g., files are okay, but DMSII
database invocations are not. PROLOG and EPILOG procedures can be
defined within the Structure Block to initialize the block's members
upon instantiation (at first reference) and wrap up prior to deallocation.

When you call a public procedure member of a Structure Block, the
addressing environment is similar to that for FIBSTACK: the members of
the block's body appear to have been declared in the procedure's
containing block, which in fact they were. Thus, they can be accessed by
the procedure using address couples.

The Structure Block type declaration generates some "stack-building"
code to provide basic initialization during instantiation. Each
Structure Block instance is represented by a data descriptor. These
instance descriptors are initialized by the compiler as a request token,
similar to the request tokens used for arrays. When one of these request
tokens is first touched, it causes a system interrupt. In response to
that interrupt, the MCP allocates space for the instance, stores in that
memory area an MSCW and a "Selector" SIRW similar to those used by
FIBSTACK, and then executes the stack-building code in the context of
that memory area as if it were a piece of a normal stack. Zeroes are
pushed by ZERO operators to initialize scalar variables, request tokens
are constructed and pushed for arrays and other complex entities, and
PCWs are pushed by MPCW operators for the public and private procedures.
Any PROLOG procedure would be executed at the end of the stack-building
code.

Public members of an instance are accessed the same way that the LIO
procedures are accessed in a FIB -- the compiler generates code to index
the instance descriptor and load word 1 (the Selector SIRW) onto the
stack. It then generates code to OR an offset for the desired member
into that copy of the SIRW, at which point you can use that SIRW to
access one of the data members or call one of the procedure members.

To illustrate, here is a very simple program using Structure Blocks:

BEGIN
TYPE STRUCTURE BLOCK DEMO;
BEGIN
PRIVATE INTEGER COUNT;
PRIVATE INTEGER VAL;
PRIVATE REAL AVG;
PUBLIC INTEGER PROCEDURE INC(I);
VALUE I; INTEGER I;
BEGIN
COUNT:= *+1;
VAL:= *+I;
AVG:= (AVG*(COUNT-1) + VAL)/COUNT;
INC:= VAL;
END INC;
PUBLIC REAL PROCEDURE GETAVG;
GETAVG:= AVG;

PROLOG PROCEDURE INIT;
BEGIN
AVG:= VAL:= COUNT:= 0;
END INIT;
END DEMO;

DEMO B;
DEMO ARRAY A[0:9];

B.INC(4);
A[3].INC(7);
END.

The code to call B.INC looks like this:

NAMC (02,0004) 5004 push SIRW to B onto the stack
ONE B1 push literal 1
INDX A6 index the instance descriptor
LOAD BD load the Selector SIRW at B[1]
LT8 6 B206 push the offset to INC
IMKS CF Insert a MSCW below the Selector
LOR 91 OR the offset into the Selector
LT8 4 B204 push the parameter (4) to INC
ENTR AB enter the INC procedure
DLET B5 pop the return value, not used

Since this is the first time that B was actively referenced in the
program, an interrupt would have occurred on (I think) the INDX operator
to allocate and initialize the instance.

The code to access A[3].INC is similar to that for B, just prefixed by
an extra step to index A to fetch the instance descriptor.

Paul

Scott Lurndal

unread,
Jun 16, 2014, 1:05:24 PM6/16/14
to
Paul Kimpel <paul....@digm.com> writes:
>On Sun, 15 Jun 2014 03:30:49 -0600, Louis Krupp wrote:

>
>The FIBSTACK mechanism that Louis mentioned was the first (that I know
>of) to exploit this addressing technique, and it's a good example to
>explore. It was implemented as part of the complete rewrite of Logical
>I/O (LIO) that was released in the Mark II.0 MCP. LIO is the user-level
>API used to access files. The original implementation was based on that
>of the B5500, but was an ambitious generalization of many of the B5500
>I/O concepts, including the idea of File Attributes. Alas, the original
>implementation was complex, buggy, and really, really slow.
>
>Part of the problem was that with all of that generalization, there were
>a lot of features and options to choose from, and if you tried to
>implement selection of the necessary features in a straightforward way,
>it took a lot of tests to figure out what a particular I/O call had to
>do. The solution was to break up that one big, mother READ routine into
>a bunch of small, tightly focused routines that each addressed a
>specific situation, and to build a mechanism that could efficiently
>select and execute the appropriate one in any given case.

We had a similar issue on Medium systems. Some of it was fixed
in MCPIX (B[234]900) and the rest was fixed with MCP/VS 2.0 by
adding so-called "tailored I/O" paths to the LIO subsystem which
would be chosen at open time.


Ivan Godard

unread,
Jun 17, 2014, 12:07:35 AM6/17/14
to
On 6/16/2014 6:01 AM, Paul Kimpel wrote:

<snip> marvellous explication of FIBSTACK omitted

> Paul

I'm guessing that you started after the move to City of Industry, which
was when I left. You mention the Hauck and Dent paper - Ben Dent was by
far the best manager I have ever worked under in my life.

I would think that the FIBSTACK trick could have been used to implement
the message type in the DCALGOL compiler that I implemented, although it
is not clear to me how the corresponding queue type would work.

Ivan Godard

Paul Kimpel

unread,
Jun 17, 2014, 11:06:40 AM6/17/14
to
-------- Original Message --------
Subject: Re: Named Memory
From: Ivan Godard <iv...@ootbcomp.com>
To:
Date: 6/16/2014 9:07 PM

> On 6/16/2014 6:01 AM, Paul Kimpel wrote:
>
> <snip> marvellous explication of FIBSTACK omitted
>
>> Paul
>
> I'm guessing that you started after the move to City of Industry, which
> was when I left. You mention the Hauck and Dent paper - Ben Dent was by
> far the best manager I have ever worked under in my life.

I started with Burroughs in mid-1970, which was indeed after the move to
Proctor Avenue, but I was just a software tech on the East Coast. I
never worked for the Engineering groups.

I have had the pleasure of corresponding with Ben Dent a couple of times
over the past year in relation to a B5500 emulation project I've been
working on. He's still a very sharp and interesting guy. He says that
he's no longer competent to write Algol programs, but I'm not sure I
believe him.

> I would think that the FIBSTACK trick could have been used to implement
> the message type in the DCALGOL compiler that I implemented, although it
> is not clear to me how the corresponding queue type would work.
>
> Ivan Godard

The technique used with FIBSTACK is a fairly general object-oriented
mechanism with nice hardware support, so it can be used anywhere you
need to maintain private state for an entity. Dave Dahm and Roy Guck
used a similar approach to implement a considerably more complex
addressing environment for the DMSII database management system.

Structure Blocks (and a related mechanism, Connection Libraries) were
implemented in the early '90s to make the technique easier to use. These
newer mechanisms are now used extensively in the system software,
especially that for networking, but they are freely available for use in
user code as well.

I would think that a FIBSTACK-like implementation would have been even
more applicable to DCALGOL Queues than Messages, but from other things
you have hinted at in earlier posts on this newsgroup, it's not clear to
me that the DCALGOL you worked on is quite the same as the one I grew up
with in the early '70s.

Paul

Scott Lurndal

unread,
Jun 17, 2014, 11:23:48 AM6/17/14
to
Paul Kimpel <paul....@digm.com> writes:
>-------- Original Message --------
>Subject: Re: Named Memory
>From: Ivan Godard <iv...@ootbcomp.com>
>To:
>Date: 6/16/2014 9:07 PM
>
>> On 6/16/2014 6:01 AM, Paul Kimpel wrote:
>>
>> <snip> marvellous explication of FIBSTACK omitted
>>
>>> Paul
>>
>> I'm guessing that you started after the move to City of Industry, which
>> was when I left. You mention the Hauck and Dent paper - Ben Dent was by
>> far the best manager I have ever worked under in my life.
>
>I started with Burroughs in mid-1970, which was indeed after the move to
>Proctor Avenue, but I was just a software tech on the East Coast. I
>never worked for the Engineering groups.
>
>I have had the pleasure of corresponding with Ben Dent a couple of times
>over the past year in relation to a B5500 emulation project I've been
>working on. He's still a very sharp and interesting guy. He says that
>he's no longer competent to write Algol programs, but I'm not sure I
>believe him.

Ben appears in a couple of the Burroughs films on youtube, the b6500 status
report and the B5500/B3500/B2500/B300 video. see channel slurn45.

scott

Roger

unread,
Sep 29, 2014, 9:43:33 AM9/29/14
to

"Paul Kimpel" <paul....@digm.com> wrote in message
news:lnmpqc$uid$1...@dont-email.me...

> Wonderful description of FIBSTACK, STRUCTURE BLOCKS Etc.

Thank you for this wonderful write-up. It deserves to be up on a web page
somewhere - not buried deep in a newsgroup like this! :-)

I too remember the upgrade from the old Logical IO procedure to the FIBSTACK
mechanism. It not only made the machine do IO so much faster but meant that
bugs in the IO routines were much more easy to understand and fix.

During that era I was working for the Large Systems Support Group in London.
(Joined Easter, 1970) We were helping with sales support and fixing the
problems being experienced on their trial machines by big customers like
Midland Bank (now HSBC), Barclays Bank, Police National Computer Unit and
one or two others.

Structure Blocks have arrived since I left Unisys and I'm glad of an
appreciation of them.

Thanks Muchly

Roger Jefferyes

P.S. Does anyone know what became of Russ Howard who was in charge of LSSG
when I arrived? Is he still with us?



One Second Burden

unread,
Jun 12, 2015, 10:09:46 PM6/12/15
to
On Monday, June 16, 2014 at 6:01:01 AM UTC-7, Paul Kimpel wrote:

> On the contrary, readers of comp.arch should be vitally interested about
> FIBSTACK, since it's all about architecture.

Hi Paul,

I stumbled on this post by accident. Wow, does it take me back! Excellent discussion of FIBSTACK and structure blocks.

And of course you're absolutely correct that Structure Blocks were a generalization of FIBSTACK that was safe for use in user programs.

Randall Gellens

unread,
Jun 15, 2015, 6:08:29 PM6/15/15
to
> And of course you're absolutely correct that Structure Blocks were a generalization of FIBSTACK that was safe for use in user programs.

By the way, Structure Blocks grew out of Connection Libraries, which were originally developed as a means of supporting protocol-agnostic networking (the ability to drop-in modules for any suite of networking protocols) and zero-buffer-copying during network operations. At the time, the fight between TCP/IP and OSI was very recent and it seemed likely that the MCP would need to support any number of networking technologies. At the same time, there was a desire to significantly speed up network I/O, and buffer copying seemed a good place to start. Connection Libraries allowed a program to link to a network support library and call a network write function which would have direct access to the data, avoiding copies on write, and allow network support library to directly call the program's input function as new data arrives, avoiding buffer copies on read. It also permitted enhanced control over linkage, so each side could control what it allowed the other side to do, and allowed bilateral unlinkage.
0 new messages