Honeywell 6000 series
The Honeywell 6000 series was a long-lived family of mainframes, in production from 1970 to 1989. They are probably best-known now for being the machines that Multics ran on for most of its life, after the initial period on a General Electric machine.
They were descendants of the GE 600 series family; after GE's computer business was sold to Honeywell in 1966, the 6000 series were Honeywell's replacements. They used TTL SSI integrated circuits and larger printed circuit boards, unlike the older (and obsolescent) discrete transistor GE machines.
The basic architecture of the GE and Honeywell series was the same: a tightly-coupled multi-processor, with all the CPUs sharing access to a collection of multi-port memory units. All used 36-bit words, and almost identical instruction sets; Honeywell added an Extended Instruction Set (EIS) which provided string-manipulation instructions, packed decimal instructions, etc.
Most 6000 series machines ran GCOS (General Comprehensive Operating System), an evolution of the earlier GECOS (General Electric Comprehensive Operating Supervisor) - early Honeywell documentation continued to call it GECOS.
Some models in the line had the additional hardware - the 'Appending Unit' (APU) - needed to implement the segmented single-level memory used by Multics, which ran only on those models of the GE 600 and Honeywell 6000 series lines.
Systems were constructed of 3 main kinds of units: CPUs, I/O controllers, and memories (technically, System Control Units, or SCUs, with the memories being an integral part of the SCU - although later documentation relegates the SCU to being a logically separate control unit, and refers to the combination of an SCU and memory as a Central Memory Unit, or CMU - although this term seems not to have caught on widely, probably because the term 'SCU' was too engrained by that point).
CPUs were connected to SCUs (with a separate cable from each CPU to each SCU); I/O controllers were also connected to SCUs (again, a separate cable for each pairing), and could send interrupts to the CPUs via the SCUs.
The first generation of I/O controllers was the IOM (Input/Output Multiplexer); these were later replaced by the IMU (Information Multiplexer Unit), which was programmable. Again, the term 'IOM' seems have persisted as the name for I/O Controllers.
All disk drives, tape drives, etc were connected to the IOMs, via mass storage and magnetic tape processors, respectively. Eventually a 'Microprogrammed Peripheral Controller' (MPC) was used to control disk and tape drives.
Serial lines, etc, were connected to a Front End Processor ('FNP', in Multics jargon), which were connected to IOMs (although some documentation indicates they could be connected directly to SCUs - although perhaps only on GCOS machines). There were several generations of these.
'Unit record' devices such as card readers, etc, were attached to a 'unit record processor', likewise attached to an IOM.
The maximum numbers of CPUs, IOMs, etc which could be connected to one system varied from generation to generation, and also between models; the GCOS-only units generally had lower limits than the Multics-capable ones.
This was in part a cost measure; the point-point cabling system required a separate port on each unit for each connection, and providing more ports required more hardware, increasing the cost. Even so, on many units, extra ports were installable options (although obviously there had to be provision for such).
Most SCUs provided 8 ports, but many CPUs and IOMs had only 4 ports, and could thus only connect to up to 4 SCUs. There were other limits, too; a CPU had a 'CPU number', assigned via switches, and many GCOS-only CPUs had only 2 bits of CPU number, whereas Multics-capable CPUs had 3.
The practical limit to the size of systems was caused by most SCUs having a maximum of 8 ports; each CPU used one port, as did each IOM. Later on an attempt was made to 'break' this limit by developing 'port sharing' technology. 6-CPU Multics systems did exist, however.
There were several generations of 6000 series machines, although some of the different names were more marketing gloss than significant changes. Performance improvements between the various generations were minimal.
Performance differences between the models of a generation were often a result of the usual marketing-driven nonsense; i.e. there was really only one kind of machine, and the lower-performing models in the range had been 'crippled' somehow (e.g. slower clocks), to reduce their performance.
Describing the various models in each generation is further complicated by the fact that model numbers were allocated to configurations; i.e. a system with one CPU would be a model xxx, and one with two CPUs would be a model yyy.
Note: The data below is generally sourced from Honeywell documentation, including marketing material. It is occasionally inconsistent as to the maximum number of CPUs, etc supported on any model. It is possible that the specifications changed over time.
The first incarnation, announced in February 1971; the CPU speed was about 1 MIPS. Model numbers of the form 60xx indicated a GCOS machine; 61xx numbers indicated Multics.
The GCOS machines were the models 6030, 6040, 6050, 6060, 6070 and 6080. The lower-priced, odd-third-digit models did not include the EIS; the others did. A low-end model, the 6025, was introduced in 1973. Systems could have a maximum of 4 CPUs and 4 IOMs.
In 1973 the model 6180, which supported Multics, was added. It was significantly different, architecturally, from the predecessor GE 645: on the GE machine, the 'protection rings' used for security on Multics had been simulated in software; on the 6180, rings were implemented in hardware. This allowed all cross-ring memory references to be checked in hardware, instead of software (the source of many security holes when the checks were not performed properly).
Series 60, Level 66 and Level 68
Announced in April 1974, introduced in 1975, these were re-badged versions of the 6000 series, in slightly lower cabinets; they did, however, offer larger memory units. The incandescent light bulbs in the control panels (which changed from the white of the original 6000's, to black) were replaced by LEDs.
The Level 66 machines were GCOS, and Level 68 were Multics.
GCOS models included the 66/05, /10, /20, /40, /60 and /80.
The Multics models were the 68/60 and 68/80, which were identical except that in the former, the cache was disabled. Systems could have a maximum of 7 CPUs and 4 IOMs, although the total of the two was restricted to 8.
Level 66/DPS and Level 68/DPS
A 1977 re-naming of the line (no hardware changes); the Level 66's were GCOS, and the Level 68's were Multics. The names DPS-68 (and presumably DPS-66, to match) were also used.
GCOS models seem to have included the 66/DPS05 and /DPS1 through /DPS5, with varying numbers of CPUs (2, 3 and 4 in the last three). The /DPS1-/DPS5 could all be configured with up to four SCUs and four IOMs.
A lightly re-engineered version (about 1/3 of the boards were identical; 1/3, or slightly more, were lightly modified; the rest were totally different) of the previous generation, announced in October 1979, and released that year.
The extensive 'lights and switches' maintenance panels of the earlier machines were replaced with a console terminal, driven by a micro-computer, the 'diagnostic processor', which interfaced to the CPUs, SCUs and IOMs; the smaller configuration panels were retained.
DPS-8 systems supported a maximum of four SCU's, down from the eight of earlier models, although each DPS-8 SCU could provide up to 16 Mbytes of memory, for 64 Mbytes total. A maximum of four IOMs were normally supported. The SCUs and IOMs were not separate cabinets, as in early models, but shared a cabinet with the CPU.
The low-end GCOS models - the DPS-8/20 and DPS-8/44 - used microcode, instead of being hard-wired (as all the other 6000 series processors were). Other GCOS models included the DPS-8/47, /49, /52, /62 and /70 (the first two being implemented in 74F technology). The /47 was limited to a single CPU and IOM, and the /49 to four CPUs and two IOMs; the /70 could support up to four IOMs.
The Multics units were the DPS-8/M models - the DPS-8/52M, DPS-8/62M and DPS-8/70M; all likewise limited to 4 SCUs (although this was not a problematic limit, as SCUs now could handle much more memory). Apparently all three used the same hardware, but the two lower-performance one has delays inserted into their clocks. The /70M came with an 8KW cache; later, an optional 32KW cache was introduced. The performance with the 8KW cache was about 1.68 times that of the 6180; with the 32KW cache, about 1.85.
NSA architectural extension
In mid-1973, a team looked at the possibility of porting GCOS to the Multics hardware, to gain increased security, but this was decided to be un-economic.
The effort did result in a new architectural extension to the 6000 Series, the 'New Series 6000 Architecture' or 'New System Architecture' (NSA); also later known as the 'Advanced Development Program', ADP, and 'Virtual Memory and Security' (VMS).
It added a set of registers used to produce a more secure system, using domain ideas and 'working sets'; and also virtual memory (which the GCOS machines had not previously supported - they only had base and bounds memory management hardware).
Implementation was delayed until the late 1970s, as GCOS was not in a position to use it until then. The concept was that the Series 60 machines could have either a Multics-style AU ('Appending Unit'), or the NSA-style 'VU' (Virtual Unit') optionally added to the CPU; the VU was actually available as an add-on option for deployed Level 66 machines.
The later DPS/8 machines also implemented as standard the 'VU' hardware, which was used by the GCOS 8 version of GCOS, which came into use starting in 1980.
- From GECOS to GCOS8: Part II - The Honeywell years
- CP-6 Internals Seminar - Covers the NSA architecture at the start
- Multics and Related 6000 Series Front Panels - Images of most of the front panels