Advanced Computing in the Age of AI | Thursday, March 28, 2024

IBM System/360: The Original Enterprise Tech 

Five decades ago, after three years of development in what was one of the most expensive and gutsy moves ever made by a public company, IBM launched the System/360 mainframe and in a way helped to create the modern computer business that we know today.

That these virtual card wallopers are still around is a testament to the fact that software is sticky, that change is difficult or sometimes not worth the trouble, either technically or economically, and that gradual evolution is what makes IT products endure. The mainframe that came out in 1964 bears some resemblance to the family of machines that bear the System/360 label, and ancient code that was written in the 1970s and 1980s still lurks in the memories and disk drives of the many thousands of very large customers who still use mainframes today. (It has been a long time since I did this, but when I used to access a vast database of 500,000 IT sites to do my statistical work for newsletters about mainframes, the database ran on a mainframe and the first thing I would see was this message as MVS booted up my 3270 green-screen session: NO RDR, NO PRT, NO PUN. That was MVS, now known as z/OS, looking for a card reader, a printer, and a punch card. Even after porting to six-core processors running at 5.5 GHz and with layers of cache and main memory and all kinds of other accelerators, the skeleton of a System/360 – and its predecessors – is still lurking inside the machine.

In December 1961, IBM commissioned the engineering study that would bring Gene Amdahl, the chief architect of the line of compatible machines, and Fred Brooks, the project leader who had to cope with the software stack and who learned the hard way that throwing people at massive software projects does not speed them up, but slows them down, together. At the time, IBM had $2.2 billion in revenues. Adjusting that for inflation today, that is on the order of $17.3 billion, and the $254 million in net income the company had prior to the mainframe would scale up to just under $2 billion in 2014 dollars. To create the System/360 line, which would take not just three years to deliver, but more like seven including the full software stack, cost IBM $5 billion in 1961 dollars, or about $39 billion today.

Show me a bet where Facebook, Google, Amazon, or any of the hyperscale startups put two years of revenue on the line for a project. And not when they were a startup with hardly any revenues, but when they were already a blue-chip, public company. By any measure, the System/360 was a bold move, and one that made IBM richer beyond its imagination. By the way, the original "Next Product Line" plan was for IBM to spend $675 million on factories, hardware development, and software development. So to say that it was slightly over-budget was an understatement. (That is again in 1961 dollars.) IBM was breaking ground in so many new technologies, from chip manufacturing to software development, that it would have been hard to keep to the schedule and within the budget. The System/360 also turned IBM into a chip manufacturer on a large scale for the first time, and it also made the disk drives and reel-to-reel tapes that are visually synonymous with the mainframe in culture.

By the fall of 1966, IBM had sold 4,000 of the mainframes and had another 20,000 on order, and it did not take long for Big Blue to make it all back. IBM grew revenues and profits by between 20 to 25 percent per year in the late 1960s, and sales flattened out in the early 1970s as early demand for mainframes peaked and it started delivering System/3 minicomputers, cutting into mainframe sales a bit. Growth resumes shortly as the mainframe had a resurgence and minicomputers took off, giving IBM two healthy – although unfortunately incompatible – product lines, which incidentally live on as the Power Systems and the System z mainframe today.

ibm-system360

The success System/360 and its descendants also eventually got it into trouble – again – with the US Department of Justice's Antitrust Division. The monopolistic practices that IBM engaged in during the punch card era were reprised during the mainframe era, and the Consent Decree that IBM signed in 1956 to get the 1952 lawsuit against the company dropped was stretched to cover the System/360 and its descendants as well as the proprietary S/3 minicomputer and its follow-ons aimed at midrange companies that was established in 1969. That was the same year that the Justice Department sued IBM again on antitrust grounds, and that lawsuit dragged on until 1982, when the government dropped the case. IBM was able to wiggle out of that Consent Decree in July 2006, removing any restrictions on Big Blue's practices. But that was only after the mainframe had taken some very serious hits as Unix and PC servers had ascended in the datacenter, and more importantly, IBM was forced to open up certain aspects of the mainframe.

There were so many innovations in the System/360 that it is hard to know where to begin to list them, but the thing to remember is that the engineering was not the only factor that was important. The Consent Decree governing IBM's behavior set a more level playing field than there might otherwise have been in data processing, as it was called then, and without that leveling, the course of information technology would no doubt be different today.

The engineering principles behind the System/360 are embedded in its name: it was meant to address the needs of companies of all sizes and all industries, and it was intended to be used for both commercial and scientific workloads. Up until this point, IBM had a slew of incompatible machines, each of which had their own research, development, and manufacturing operations. This was annoying and costly for customers, so the idea with the System/360 was to make a single line, spanning a performance range of 50 to 1 on a relative performance scale (measured in millions of instructions per second, or MIPS), that could all run the same operating system and application software. This seems obvious now, but it was not clear how to make a general-purpose computer back then that could do so many different things. There were only around 35,000 computers in the world back in 1965, when System/360 sales started taking off and the projection was for there to be maybe 85,000 a decade later.

Both Tom Watson Senior, who ruled IBM during the punch  card era, and Tom Watson Junior, his son who brought the System/360 to market and commercialized computing on a scale never seen before, had a habit of underestimating the demand for these devices.

The System/360 processors pushed the limits of the circuitry of the time, cramming an order of magnitude more electronics into the same space as could be done with prior systems based on vacuum tubes. Performance on the top-end System/360-75 was about 150 times greater than that tube-based system, and the cost for a unit of work went down by a factor of 40. IBM did something else radical: It made the entire line available at the same time, and it created a standard peripheral interface between the central processor complex – the main frame – and its storage and end user devices. The System/360 could do decimal math as well as floating point math, which made it useful for all kinds of calculations. And a little more than a decade later, when Seymour Cray created the vector supercomputers that bore his name and are synonymous with supercomputing, IBM would strap vector coprocessors, called the Vector Facility, onto mainframes to create the world's first ceepie-geepie systems to compete. To one way of thinking, the mainframe is an architecture that constantly absorbs new accelerators and technologies to justify the high price IBM commands for these systems.

ibm-tom-watson-system360

Tom Watson, Jr and the System/360

The Consent Decree, among other things, compelled IBM to publish the specifications of its System/360 machines and to allow for third party equipment makers to create alternatives to its products and, importantly, set the precedent for using them without breaking warranty agreements with IBM. IBM also had to provide operating systems to third party clone mainframe makers. The Decree also compelled IBM to offer machines for sale that were reasonably close to the rental price, helping to foster the third party equipment leasing business, and also made IBM publish list prices for all system components, so everyone knew where the ceiling was from which to negotiate down to the street price – including competitors. By the early 1980s, these practices were so ingrained that the IBM PC was effectively created as an open system, with Intel as the chip supplier and Microsoft as the operating system supplier, setting the stage for the next computing era.

Back in the mid-1990s, IBM moved away from expensive bi-polar chip technology, which ran so hot that it required water cooling and was also expensive, and toward the CMOS processes that other chips employed. And since that time, IBM has done what other chip makers have done: add caches, crank the clocks, add cores, and boost the thread count and the memory capacity to keep the performance of the systems growing.

IBM still invests in mainframe hardware, but these days the Power Systems and System z families share a lot of peripherals and, in fact, whole chunks of their processors are carved from the same basic transistor blocks. But they remain distinct. The mainframe was the first platform to get logical partitioning (Clone system maker Amdahl did it first, and IBM copied it), and that helped drive up utilization on systems to 95 percent and above. The mainframe started out in the System/360 era as a batch processor with high I/O throughput, and to this day, even though it runs some of the largest databases in the world, the mainframe is best fit for workloads with high I/O bandwidth needs. The machine pays for itself because it is incredibly rugged and can be run at very high utilization, day in and day out.

The System z line is still generating around $5 billion a year in hardware and base operating system sales. (It was north of $13 billion at the peak of the market in 1990, just for comparison.) The top companies in the world still deploy mainframes:  23 of the top 25 banks, 10 of the 10 top insurers, 23 of the top 25 retailers, and hundreds of governmental bodies rely on mainframes. By some estimates, there has been over $1 trillion in software investments in mainframe software in the past five decades, 70 percent of the data used by large enterprises resides on storage attached to mainframes, and 60 percent of the transactions run in the world touch a mainframe at some point. Of the companies in the Fortune 500, 355 of them have mainframes – still. And heaven knows many have tried to sell replacements to them.

On top of that System z hardware, there is a stack of systems software, including file systems, databases, middleware, and development tools, that brings in several billions of dollars more on top of that, and maintenance on the whole shebang rakes in even more money. A decade ago, IBM rejuvenated the System/390 mainframe business by porting Linux to the machine, by offering specialty configurations of the mainframe engines that were cheaper and designed to accelerate database, Java, and XML workloads, and by shifting to utility-based pricing for its software stack. And even though the mainframe base has shrunken to somewhere around 6,000 customers or so, these cheaper MIPS for Linux and acceleration of software functions have kept the installed base of MIPS growing at a pretty steady clip, year in and year out. That keeps the software revenues growing, and this is what IBM really cares about.

Expanding the portfolio of applications that can run on the mainframe is what has kept the platform alive for 50 years. It is safe to say that without the addition of Java and WebSphere in the mid-1990s and Linux to the System/390 in 1999, mainframe processing capacity would have dropped significantly, and the other specialty engines have helped keep the capacity being turned on inside the frames even as the price has come down. Some would say the adoption has been enabled because the price has come down. Speaking at the mainframe 50th birthday celebration at the Walldorf Astoria hotel in New York on Tuesday, Steve Mills, who is in charge of both Software Group and Systems and Technology Group, said that MIPS shipments have grown by double digits in the past decade and the aggregate number of MIPS in the field is 5X that of a decade ago. With mainframe revenues coming down a little bit over that term, that means prices have come down even faster than the base is growing, and this is how IBM has been able to attract new workloads to the venerable mainframe.

At the 50th birthday celebration in New York today, IBM is rolling out new mainframe software for new workloads.

The first is called zDoop, and it is a variant of IBM's BigInsights variant of Apache Hadoop that is running on Linux instances on the mainframe. IBM has partnered with Veristorm to integrate VSAM flat files, logs, DB2 databases, and other system files on z/OS to the HDFS file system underneath Hadoop. The idea is to keep the data on the mainframe where it is safe and secure behind the mainframe's legendary RACF security and authentication software.

IBM is also rolling out variants of the System z EC12 high-end and BC12 midrange mainframes that turn them into cloud platforms. The machines are equipped with either DS8000 or V7000 storage arrays, all tooled up with Linux instances and SmartCloud Orchestrator (an OpenStack variant), OmegaMon XE, and Tivoli Storage Manager. The Enterprise Cloud System can scale up to 6,000 virtual machines in a single EC12, and will come with special utility pricing for managed service providers so they can get started without having to spend millions of dollars up front for the stack. IBM says that the Enterprise Cloud System can offer a 55 percent lower total cost of ownership compared to clusters of X86 iron.

This is the kind of sales pitch that enterprise customers have heard for five decades now, and will probably hear for a few decades more.

EnterpriseAI