The History of the HP 3000

by Bob Green

In the 80's, HP introduced a new PA-RISC hardware architecture for the HP 3000. The original 16-bit architecture became known as the Classic. HP continued to produce it for many years. The operating system that the Classic runs is MPE V. Although PA-RISC is drastically different in technology, its MPE/iX operating system can run most programs from the Classic HP 3000 in Compatibility Mode.

On November 14, 2001, HP announced the sunset of the 3000 line, with HP support ending on December 31, 2006.

History of the Original HP 3000

Did you know that the first HP 3000 was a complete disaster and was withdrawn from the market for a short time? Did you know that Hewlett-Packard originally planned on a 32-bit machine, not a 16-bit one?

Background (1964-1969)

This product was incredibly successful, especially with schools. The original 2000A system was created by two guys working in a corner: Mike Green who went on to found Tandem much later and Steve Porter who also went on to found his own computer company. Heavy sales of the 2000 brought the computer division of HP its first positive cash flow, and with it the urge to "make a contribution." The engineers and programmers in Cupertino said to themselves: "If we can produce a time-sharing system this good using a junky computer like the 2116, think what we could accomplish if we designed our own computer."

Abortive First Try (1969-1970)

The Omega was designed to compete with big CPUs. But Omega looked too risky to management. HP would have had to borrow long-term funds to finance the lease of machines to compete directly with IBM. So it was cancelled. Some of the Omega architects left HP, but most stayed. "Several people who remained took to wearing black-velvet armbands, in mourning for the cancelled project," according to Dave Packard in his 1995 book, The HP Way.

The 16-Bit Alpha (1970-71)

Same People, Smaller Hardware, Bigger Software

To make the Alpha fit within budget and microcode restrictions, data segmentation had to be dropped. There wasn't room for both it and floating-point arithmetic. As a compromise, the expandable DL area of the stack was added. The equator of the user data stack was the DB register; the Z register defined the maximum DB+ address and the DL register defined the maximum DB- address. This strange feature, which provides for both positive and negative data addresses, is the one serious design error in the HP 3000. It made byte addressing ambiguous and led to numerous bugs in the MPE code.

The instruction set of the Alpha was designed by the systems programmers who were going to write the compilers and operating system for the machine. The prevailing "computer science" philosophy of the day was that if the machine architecture was close to the structure of the systems programming languages, it would be easier to produce efficient, reliable software for the machine and you wouldn't need to use Assembler (that is, a high-level language would be just as efficient and the code would be much easier to maintain).

The Alpha was a radical machine and it generated infectious enthusiasm. It had virtual memory, recursion, SPL instead of Assembler, friendly MPE with consistent batch and on-line instead of OS-360 with its obscure command syntax, variable-length segments instead of inflexible pages, and stacks instead of registers. The Alpha was announced as the HP 3000 with a fancy cabinet of pizza-oven doors, available in four colors. Prospective users were assured that it would support 64 users in 128 KB of memory.

Harsh Realities (1972-73)

I worked at Cupertino at the time and was assigned to coordinate the production of the ERS (External Reference Specifications) for the new software. I was as excited as everyone else. The first inkling I had that the HP 3000 was in trouble came in an MPE design meeting to review the system tables needed in main memory. Each of the ten project members described his part of MPE and his tables: code segment table, data segment table, file control blocks, etc. Some tables were memory-resident and some were swappable. When the total memory-resident requirements were calculated, they totalled more than the 128 KB maximum size of the machine.

MPE wouldn't fit, so everyone squeezed: the programmers squeezed in 18-hour days, 7 days a week trying to get MPE to work. Managers were telling their bosses that there was no problem, they just hadn't had a chance to "optimize" MPE yet. When they did, the managers maintained, it would all turn out as originally promised. So marketing went on selling the machines to the many existing happy users of the HP 2000. As the scheduled date for the first shipment approached, the Cupertino factory was festooned with banners proclaiming "November Is a Happening."

The first HP 3000 was shipped November 1, 1972 to Lawrence Hall of Science in Berkeley, California. But it was incomplete: it had no spooling, no real-time, etc. It supported only two users, and it crashed every 10 to 20 minutes. Customers who had been promised 64 terminals and who were used to the traditional HP reliability became increasingly frustrated and angry.

Eventually the differences between the HP 3000 reality and the HP 3000 fantasy became so large and well-known that there was even a news item in ComputerWorld about it -- the first bad press ever for HP. Bill and Dave were not amused. The product was withdrawn from the market for a short time.

Struggling to Restore Lost Credibility (1973-74)

Once the HP managers realized the magnitude of the 3000 disaster, the division was in for lean times. Budgets and staffs that had swollen to handle vast projected sales were cut to the bone. Training, where I worked, was cut from 70 people to fewer than 20 in one day. HP adopted a firm "no futures" policy in answering customer questions (a policy that lasted for years after the HP 3000 trauma, but was forgotten by the time of the Spectrum-RISC project). The new division manager was strictly no nonsense. Many people had gotten in the habit of taking their coffee breaks in the final-assembly area, and kibitzing with the teams testing the new 3000s. Ely banned coffee cups from the factory floor and instituted rigorous management controls over the prima donnas of the computer group.

By continuing to work long weeks, the programmers managed to reduce MPE crashes from 48 a day to two, and to increase users from two to eight. Marketing finally took a look at what the 3000 could actually do, and found a market for it as a replacement for the IBM 1130. HP no longer sold the 3000 as a souped-up version of 2000 time-sharing.

Instead they sold the 3000 as a machine with more software capability than an IBM 1130 that could be available to a number of users at once instead of just one. Eventually the 3000 became a stable, useful product.

The 24-Hour Time Bomb

To my mind, this happened when someone discovered the "24-day time bomb" bug. If you kept your HP 3000 running continuously for 24 days (2^31 milliseconds) without a shutdown or a crash, the internal clock register would overflow and the system would suddenly be set back by 25 days!

From:    Ted Workman 
To:      Bob Green
Subject: HP 3000 and MPE stability
Date:    Mon, 30 Jun 2003 03:11:54 -0400

I read with interest your web article on MPE. In particular, you wrote:

Eventually the 3000 became a stable, useful product. To my mind, this happened when someone discovered the "24-day time bomb" bug.

I was the Product Specialist for MPE for the Southern Region of the US in the mid 70s. I basically supported MPE from the hardware support organization. There were essentially one of us in each of the four regions in the US.

We had an installed base of about 20 machines. MTTF had begun to improve from hours to days. Two of my sites had machines that had settled in, and were particularly stable. We probably had 15 customer engineers for the HP3000 in the Southern Region, and they were not idle fellows.

Larry Rea, one of the best CEs we had in the region, reported that two of his sites were experiencing problems with the "Clock" card, which also controlled the console of the HP3000. These were the Series CX machines, and were pretty stable from a hardware point of view. When the same machine failed with a replaced clock card, after running for over a month. Larry contacted me for help.

At that time, MPE stored the date of startup and a double word value. milliseconds since midnight of date started up, that was updated every millisecond when the clock interrupt occurred. The variable was a 32 bit unsigned value, so when it was updated from FFFF FFFF, it rolled over to 0000 0000. When MPE displayed the current date and time, the calendar retreated. I had found this by reading the listings for MPE, at that time the product specialists had full access to source code, and received upgrades in the form of modification files that could be used to compile against the base source code for each module.

I instructed the clients to do a cool start once a week, and grimaced when/if they asked me why. I cannot remember what I told them, but it was not: "the original designers of MPE never thought the OS would stay up for 25+ days in a row"!!!

When I left the Houston field office, I went to HP Cupertino and started up the Performance Engineering group in the Marketing Support group, working for Roy Clifton.

I enjoyed your article!! I now work as an Attorney, specializing in Intellectual Property law.

Ted Workman, Taunton, MA

The Comeback: Fulfilling the Promise (1975-76)

The Series II made a single change that was not backward compatible: Long floating-point was expanded from 3 to 4 words. This caused the Chronos intrinsic to be dropped, replaced by the Clock and Calendar intrinsics. Chronos was the only MPE intrinsic that I had designed. With that exception, the upgrade from the Series I to the Series II was simple and painless.

The choice of SPL as the machine language instead of Assembler truly began to pay off now in an avalanche of excellent software: the IMAGE database (again, two guys working in a corner: Jon Bale and Fred White) was soon joined by compilers for COBOL and RPG, a screen handler, and other tools to support transaction processing.

Concurrent, consistent batch and time-sharing was now a reality and the goal of concurrent real-time was finally dropped as unrealistic. The HP 3000 hardware now matched the software written for it. Business users discovered that the 3000 was great for on-line transaction processing; they dragged Hewlett-Packard firmly into the commercial information processing world.

The years that followed were merely an unfolding of the original MPE specifications. MPE C, II, III, IV, and V were all attempts to complete what had been promised to the first users. The development of the 3000 was an evolution, not a revolution. At last, with the Series 64 in 1982 the 3000 reached the 1972 target of 64 users on a single machine. Many of the key contributions to this history were small bootleg projects. Disc caching, for example, was again the work of two guys in a corner.

What Has Changed About the HP 3000?

What Has Not Changed?

For another interesting history of the HP 3000, read

HP's EARLY COMPUTERS, Part Three: THE STRONGEST CASTLE: The Rise, Fall and Rise of the HP 3000 by Chris Edlar.

Reader Feedback:

Just read your excellent brief History of the HP 3000. We got the first HP 3000 in Washington State. A lot of the error searching was done in the R. W. Beck Computer Center in Seattle on Friday night by guys who came up from Cupertino.

I saved all the programs that made the 3000 crash and the guys would come and put the patches on from the previous weeks discovery. Then I would run my crash programs, all in Fortran and the system would crash. Hours later the guys would leave with some dump tapes.

We never knew about the "24 hour bug" as our computer was always being pushed to the limit and crashed several times a day. Then we got a night operator and did a backup and reload every night.

Because HP never improved the Fortran capability, as often promised, I was finally forced to order a Honeywell DPS 8. I left that firm in 89 and went fishing.

Best Regards,

Larry Mahoney

More Reader Feedback

I read with interest your article about the history of the 3000. Of particular interest was the part that said "The Alpha was announced as the HP 3000 with a fancy cabinet of pizza-oven doors, available in four colors.".

I once saw a HP 3000 which was multiple cabinets, each of which had many wood-grain doors which were the 19" width of the rack by perhaps 10" high, each of which had an Ace lock (round key, like on vending machines). I be- lieve that at least some of the compartments behind these doors held wire- wrap boards, not etched PCBs. The "console" was a table with a pair of CRTs in the middle, and to the left and right of these CRTs were large panels full of LEDs, with stenciled-and-varnished designations for each LED. I was told that one CRT was reserved for maintenance functions, while the other was the OS operator's console.

Does this ring any bells? Was this a prototype? It doesn't sound anything like any other description of a 3000 that I've read.

This would have been in 1974 or 1975..."

Thanks, Terry Kennedy (www.tmk.com)

Note: What you are describing is the original HP 3000- the one that was put on hold and replaced 18 months later with the "CX 3000" (which did away with the fancy cabinets and desk to save money).

Woodgrain was one of the 4 colors that you could order for the original 3000, and the panels of flashing lights were quite impressive! The other colors were sea gold, marine blue and red!

Deja Vue - The Spectrum Project

By Bob Green

When we wrote our book "HP 3000 Evolution" in 2003, we included Chris Edler's history of the Classic 3000. After the initial development, the HP 3000 grew and prospered. From 1974 to 1984, HP continued to produce more powerful 3000 hardware running more capable software. Each new model was compatible with the previous version and a joy to install.

But the pressure was on to switch to a 32-bit architecture, as other manufacturers were doing. So HP announced a radical change: a new 32-bit hardware for the 3000. The project was code-named Spectrum. As a 3000 consumer and 3000 vendor, Robelle was excited and concerned about the prospect of a new hardware architecture. Certainly it would be wonderful to have more powerful processors, but what about the disruption to our steady incremental, risk-less progress.

The first notice we took of the Spectrum appeared in Robelle's December 1984 customer newsletter, with continuing news to follow for the next four years (my retrospective comments are interpolated in indented paragraphs):

December 12, 1984 Newsletter:
The first Spectrum machine will be an upgrade for the Series 68. Other size models will follow soon after, since HP is working on different Spectrum cpus in three divisions at once (in the past, all 3000 cpus came out of one division). This first Spectrum can be expected in the first half of 1986.

Please make a note of that 1986 promised delivery date and remember that HP faced serious competition from DEC and others. Customers who loved the 3000, but had outgrown the power of the system, were demanding more capable models.

The actual number of instructions in the Spectrum turned out to be about 130, not 16, but they were all simple enough to run in a single clock cycle. HP was the first computer company to go with the RISC philosophy and the only major one to risk the firm by converting all their computer models, both technical and commercial, to a single RISC design.

The expectations were building up in our minds: these machines would be much faster than our current models!

Vision was an alternative 32-bit computer project at HP, using traditional technology, which was cancelled to make way for the RISC design from HP Labs. Invoking Vision re-assured us that this project is possible, that progress is being made. It is now 6 months after the first announcement of the project.

"Most of what is printed about Spectrum is not to be trusted. Spectrum will be introduced at the end of 1985 and delivered in Spring 1986. There are 40-50 prototypes running in the lab and the project team consists of 700-800 engineers. HP will co-introduce a commercial version and a technical version with the commercial version fine-tuned to handle many interactive users, transaction processing, IMAGE access, and the technical version will be structured for computational programs, engineering applications, and factory automation. HP will eventually offer a choice of MPE and UNIX. Most software will be available on Spectrum at introduction time and over time all software will be available."

HP tried to dispel rumors, but still predicted 1986 for delivery. HP will produce two Spectrum lines: the UNIX line for technical users and the MPE line for commercial users, using the exact same hardware.

The new relational HPIMAGE database mentioned here was cancelled much later in the project, after a brief encounter with end-users. I don't remember much about HPIMAGE, except that a lot of work went into it and it didn't succeed as hoped. TurboIMAGE ended up as the database of choice on the Spectrum. Without any inside information, but based just on past experience and common sense, Robelle tried to inject some caution about the 1986 release date. During the original traumatic HP 3000 project, Dave Packard "sent a memo to the HP 3000 team. It was only two lines long and said, essentially, that they would never again announce a product that did not then currently meet specifications." [Chris Edler] The division listened for over 10 years, but eventually, people forget….

From an HP press release: "100 prototype models were already being used internally for system development as of April 1985."

HPE, the new operating system for the commercial Spectrum is a superset of MPE. It will have two modes of operation: Execute mode (HP 3000) and Native Mode. The switch between the two will be made on a procedure call, but there will be some programming work needed to translate parameters when switching.

Execute mode was eventually called Compatibility Mode and switching between modes turned out to be major cpu bottleneck in the new system, albeit one that would be removed over time.

December 30, 1985 Newsletter:
From Gerry Wade of HP: The name of the Spectrum machine, when it comes out, will not be Spectrum. Another company already has that name. Spectrum will use the IEEE standard for floating-point arithmetic and will also support the HP 3000 floating point. Each data file will have a flag attached to it that tells which type of floating-point data it contains (the formats are not the same).

The file flag idea never happened, although the TurboIMAGE database did introduce a new data type to distinguish IEEE floating point. Information on implementation details is starting to flow, which helps us believe that the project is on schedule and likely to deliver the more powerful servers we desire..

Joel was the HP Labs leader who was the spark plug of the RISC project, building on research that he had done previously at IBM. In retrospect, we can see that Joel was talking about the performance and delivery of the UNIX Spectrum, not the MPE version, but customers took this as a promise of vast performance improvements in the very near future. It is now past Spring 1986 and the promised new 3000 machines are nowhere in sight. In fact, HP has not yet announced the new models and pricing. This is the first slippage in the project, barely noticed at the time.

Since Suprtool and Qedit were essential to the large 3000 customers that HP was targeting, HP asked Robelle to start porting and testing our products on the new systems. But to do that, we had to sign a Non-Disclosure Agreement, the most draconian one we had ever seen. We used careful wording in our announcement above. From this date on, until years later, we could not tell our customers anything useful about the new machines. HP was especially sensitive about their reliability and performance.

When we arrived in Cupertino to do our first testing, we found the prototype Spectrum systems crashing every few minutes and running slower than our tiny system 37. We were appalled. Nothing in HP's public statements had prepared us for the state of the project. I had personally gone through a similar situation with the original 3000 in 1972-74, and I wondered if upper management at HP knew how terrible things were. I thought about talking to them, but our NDA also prohibited us from talking to anyone at HP.

The UNIX versions of Spectrum, on the other hand, seemed to be humming along nicely, showing that it was not a hardware problem.

September 11, 1987 Newsletter (over a year later):
First Spectrum Shipments: Rumor has it that HP shipped the first four 930 machines on Friday, August 21st, with more to follow every week thereafter. As of press time, we have been unable to find out whether ordinary mortals are allowed to touch these machines (as opposed to those who have signed non-disclosure agreements).

Due to the NDA, over a year passed with no Spectrum news in our newsletter. The project was now 18 months past the original promised delivery date, but was still far from completion. Many people wrote articles, about the Spectrum, mostly based on marketing hype from HP, but no one broke the embargo on real information. We were all terrified. The MPE group had dug themselves into a very deep hole, and no one wanted to be the one who caught the eventual backlash.

If it takes forever, we will wait for you!

For a thousand summers, we will wait for you!

'Til you're truly ready, 'til we're using you!

'Til we see you here, out in our shops!

From the HP Management Roundtable:

Schedule for Shipping Spectrums
"We are shipping equally around the world. Our first shipments went to both North America and Europe. We are acknowledging availability for 930s and 950s through December at this time … We expect by the end of November to be able to have acknowledged the entire backlog."

HP continued to spin the "shipments" of Spectrums, without mentioning that these were not finished products. The models were 930 and 950 and the operating system was called MPE/XL, changed in later years to MPE/iX when POSIX was integrated into MPE. By this time, HP was worried about their stock price also and did not want any negative news in the financial press, no matter how accurate. As shown by the next Q&A at the roundtable…

Question:
"Are the 930s and 950s being shipped or not? In public you tell me they are shipped. In private, however, I hear from both users and HP that these machines are still covered by non-disclosure agreements and that access to these machines is very restricted, even when in customer sites. What is the story?"

Answer:
"MPE/XL architecture is very, very new. There's a million new lines [of code] that go into MPE/XL, and a lot of software sub-systems as well. And so we are being extremely cautious in how we proceed at this point. We're going through what we call a slow ramp-up through the remainder of this year and going into large shipments in 1988. The reason for that is that we want to fully test out the system capability in a large number of customer environments and we want to make sure that the information on what's going on in there and the people involved are protected from outside folks who either benevolently or not benevolently would like to find out what's going on. I'm sure we're going to run into some problems along the way that haven't been encountered in our earlier phases of testing. We haven't really hit these machines with full production pressure yet. We know from experience that when you do that, you uncover things that you could never uncover in testing, even though extremely rigorous. [Rumor has it that the customers receiving Spectrums now are not allowed to put them into production until 1988.] … A lot of people out there would like to see us have trouble and we want to make sure that, if we do have some problems, they're contained in our user community and they don't get out into the press and create problems that will come back to haunt many of you".

Early Spectrum customers called us to ask which version of Suprtool and Qedit they needed for their new systems, and whether there were any problems that they should be aware of. But legally, we could not even admit that we knew of the existence of the new servers. Anyone who broke the NDA could expect to be figuratively "shot". So we came up with the following wording: "If you had a new 3000, and we are not admitting that we know anything about a new 3000, you should be using Suprtool version 3.0 and Qedit version 3.6. On this hypothetical system, it might not be a good idea to hit Control-Y while copying a file from any other HP 3000. We can't tell you what will happen, but you won't like it."

Interestingly, we didn't hear about the Manufacturing Release (MR) of the Spectrum from Hewlett-Packard itself. As far as we can determine, HP kept this event very quiet - no press conferences or splashes of publicity. Even some HP people in Cupertino were not aware that MR had occurred. Just because the 930 and 950 are released does not automatically guarantee that you can get one. Given the huge backlog of orders that HP has, it will make "controlled shipments" for a while, picking sites whose expectations match the state of the machines.

Users had been following Spectrum for almost 4 years now and you can see that we are eager for the release of the product. The MPE lab has grown to hundreds of engineers and technicians and hundreds of Spectrum servers. The amount of money being plowed into the project is awesome. Anyone with any kind of skills is being hired as a consultant, in an attempt to get the situation under control and begin shipping revenue-generating servers. But we were premature in our February proclamation of Manufacturing Release, which is an HP corporate milestone that requires signed confirmation that the product passes the performance tests set for it in the design specifications.

We also concluded, and reported, that MPE XL had obtained MR (Manufacturing Release). That is untrue. MPE XL has apparently obtained SR (System Release), but not MR. "System Release" seems to be a new category of release, created just for MPE XL. We have heard from some new 950 customers who did not need to sign a non-disclosure agreement. However, one customer reported that before HP would allow him to order, he had to sign a document stating that he had no specific performance expectations (!?). On the other hand, we heard from a school that recently went live with 35 student sessions and had great response times ("the machine is coasting along at 10% cpu utilization").

In order to stem the rising tide of bad expectations, HP released the MPE systems even though they could not pass the testing department. And the performance was still poor in many cases, less than the non-RISC 3000s being replaced, although excellent in a few other cases.

Non-disclosure continues - HP is picking their initial customers carefully and coaching them to only pass on the good news about their new systems. We are still frustrated to not be able to pass on our ideas about how users can improve the performance of the Spectrum.

October 12, 1988 Newsletter:
Childcraft / Series 950. Gary Porto at Childcraft reports that with MPE XL 1.1 the problem of a serial task in a batch job hogging the system is not so bad as it was with 1.0. This problem can occur with SUPRTOOL, QUERY, or any long serial task. The batch job still hogs the system, but at least people can get a minimum of work done. With 1.0, they couldn't even get a colon! Gary reports that he has 65 on-line users on his 64-megabyte Series 950 and that the performance is pretty good - as good as his Series 70.

On the 4 year anniversary of the project, HP released version 1.1 of MPE/XL, which made the systems much more useful, but still not up to the original promised performance of 1984. However, the promise of the "Precision Architecture" (HPPA) was there, as certain tasks were amazingly fast.

By this time, HP salesmen were getting irritated with us for not giving our customers any kind of endorsement for the switch to the 930/950. But our NDA was not cancelled until Manufacturing Release. Finally, the sales force convinced HP Cupertino to send us a signed release from our NDA. I don't know when MR eventually happened.

Keith Howard of Collier-Jackson in Tampa, Florida participated in the Spectrum beta testing and upgraded from a Series 58 to a Series 950 - quite a leap. One application was found to be 6% slower due to constant switching between compatibility and native modes, but in most circumstances the machine was five to ten times faster than the Series 52 and one batch job ran 53 times faster!

Glaxo Export has temporarily deferred delivery on its second and third 950 systems due to implementation problems on the initial machine. John Walsh of Glaxo related at the Scarborough conference that the first processor was due to go live earlier this year, but when the company came to do a volume throughput test, the machine was unable to load the full operational workload of the Glaxo Series 70. ... Glaxo's problem was caused by the MPE XL 1.0 Loader table being too small, but Glaxo now has a copy of version 1.1 and is about to re-run the volume throughput test.

HP promises performance improvement for Precision Architecture over the next five years of 40-50% per year. Some of this will be achieved by further tuning of MPE XL -- version 1.1 is said to be at least 20% faster overall.

As with the original 3000 project, the birth of the Spectrum was traumatic, expensive and embarrassing. But it paid off. HP was able to roll out better servers for the 3000 line on a regular basis for the next 10 years.

Despite the numerous expansions and revisions to the HP 3000 hardware and software, upgrades have been painless. Even the conversion to the PA-RISC design was backward-compatible and reasonably painless (if you ignore the slipped schedules). Often the user just rolled the new system in on a Sunday, plugged it into the power, reloaded the files, and resumed production. The original 1974 MPE Pocket Reference Card is still useful; everything on it works under MPE/iX version 7.5, except the Cold Load instructions. I have programs that I wrote in 1972, for which the source code was lost years ago, and they still run in Compatibility Mode.

When asked for an obituary for the HP 3000, my reply was:
A great IT platform: reliable, affordable, flexible, easy to operate, and easy to program. And every release compatible with the previous for over 30 years. Perhaps some future OS team will adopt these same goals.