mainframe

mainframe n.

Term originally referring to the cabinet containing the central processor unit or `main frame' of a room-filling Stone Age batch machine. After the emergence of smaller `minicomputer' designs in the early 1970s, the traditional big iron machines were described as `mainframe computers' and eventually just as mainframes. The term carries the connotation of a machine designed for batch rather than interactive use, though possibly with an interactive timesharing operating system retrofitted onto it; it is especially used of machines built by IBM, Unisys, and the other great dinosaurs surviving from computing's Stone Age.

It has been common wisdom among hackers since the late 1980s that the mainframe architectural tradition is essentially dead (outside of the tiny market for number-crunching supercomputers (see cray)), having been swamped by the recent huge advances in IC technology and low-cost personal computing. The wave of failures, takeovers, and mergers among traditional mainframe makers in the early 1990s bore this out. The biggest mainframer of all, IBM, was compelled to re-invent itself as a huge systems-consulting house. (See dinosaurs mating and killer micro).

However, in yet another instance of the cycle of reincarnation, the port of Linux to the IBM S/390 architecture in 1999 - assisted by IBM - produced a resurgence of interest in mainframe computing as a way of providing huge quanitities of easily maintainable, reliable virtual Linux servers, saving IBM's mainframe division from almost certain extinction.

--The Jargon File version 4.3.1, ed. ESR, autonoded by rescdsk.

The defining feature of a mainframe is summed up in the word availability. This encompasses uptime, capacity for multitasking, capacity to serve many users at once, ability to service multiple large I/O requests at once, backwards-compatibility, and hardware and software reliability.

In fact, just about the only thing it doesn’t stress is raw computational power. If you need a lot of flops or a lot of parallel processing, you're looking for either a supercomputer or a cluster of some kind. A mainframe would just slow you down.

A mainframe does what it does using many technologies, some of which are being reinvented in desktop computers decades after their introduction among mainframes. The most interesting of these reinventions is virtualization, the process of running multiple guest OSes at the same time under a virtual machine. IBM was doing this decades ago with VM/370 and VM/CMS (that is, running the Conversational Monitor System as a guest under VM, their imaginatively-named virtual machine). This serves two major goals: It makes OS crashes easily survivable and debugable, and it allows very old OSes to run unmodified on modern hardware invented decades after they essentially stopped changing. (Only the virtual machine needs to know much about the hardware, because everything else has to go through it anyway.)

There have been more subtle reinventions over the decades. Web forms could be seen as a return to the block mode style of terminal interaction (as seen in the 3270 and the 5250), as opposed to the character-oriented style used in every non-IBM context to have terminals at all (as epitomized by the VT100 and its uncountable clones). The block mode style allows one server to control many more clients because all I/O is batched to be sent a screen at a time instead of being sent one line or even one key press at a time. Even with AJAX, the fundamental I/O model of Web sites looks more like block mode than anything else.

The lesson here is that the advance of technology doesn't shove the obsolete technologies completely out of the way, it relegates them to where they actually make sense. Technological advance is all about having options, which includes the option to use the “obsolete”.

A mainframe, these days, is a large data-processing computer that mostly runs legacy applications. The dominant mainframe vendor in the US is IBM, and people often believe that IBM is the only one left. IBM's mainframes are the fastest, but not close to the only ones around. They have about 90% of the US mainframe market, but in other markets, half or less of mainframe sales are IBM. Some mainframe systems vendors focus on solid transaction-processing performance, while other vendors are mostly focusing on providing continuing support to legacy users without impressive performance increases. Combined, world mainframe sales are several billion dollars per year.

Unisys is the secondary player in the US market, and leads in market share in other areas, like Taiwan and Latin America. Formed in the 80's by the merger of two earlier companies, Sperry and Burroughs, Unisys has two operating systems: OS 2200 and MCP. OS 2200 is based on the UNIVAC Series 1100, a relatively late (early 1970's) system with a modern command-line interface vaguely reminiscent of VMS. It runs on the Clearpath Dorado line of 36-bit CISC machines. MCP is arguably the oldest operating system still in use, with its first version released in 1961. It runs on the Clearpath Libra, formerly called the Burroughs Large System, a strange 48-bit stack-based architecture. Unisys is one or the few mainframe vendors left that is actively seeking new customers and developing high-performance systems.

Fujitsu pulled out of the US and Japanese markets around 2000, having previously made IBM 390 clones running MVS. They're still active in the European market with the ICL VME and Fujitsu-Siemens BS2000 operating systems, both of which have been acquired from purchasing other companies. VME has substantial, although fading, market share in the UK. In the last few years, Fujitsu has been de-emphasizing it, putting OS development on the back burner and pushing a software emulator for Windows instead of delivering new hardware. BS2000, with a large customer base in Germany, still gets new hardware and OS releases, although Fujitsu is phasing out high-end 390-based processors in favor of commodity x86 chips running a mixture of ported and emulated software.

Groupe Bull, a French vendor, produces two distinct operating systems, confusingly called GCOS 7 and GCOS 8. GCOS 8 is the older of the two, originally developed at GE as GECOS and sold to Honeywell, who sold all their OS property to Bull in the late 80's. GCOS 7 is a more recent operating system, loosely derived by Multics and developed by Honeywell in the late 70's. Both systems are getting software and hardware refreshes, with the hardware based on off-the-shelf Itanium and x86 processors with custom firmware. GCOS is used heavily in France, although its market share is negligible pretty much everywhere else.

Hitachi develops two operating systems, VOS1 and VOS3, purely for the Japanese market. VOS3 is a direct clone of MVS, and runs on 390-compatible systems called the AP8800. VOS1 is being phased out, but before the hardware's unfortunate demise ran on a POWER-derived machine (probably a rebranded 595) called the AP7000.

NEC has several high-end in-house operating systems, including a high-end UNIX and the ACOS family, three GCOS-based mainframe operating systems. ACOS-6 was a rebranded GCOS 8, and has been end-of-lifed in the last couple of years. ACOS-2, the low-end system, was based on the long-dead GCOS 4 system by Bull; you can still buy ACOS-2, if you don't mind hardware running on Pentium III Xeons with a maximum of 256MB of RAM. ACOS-4, a GCOS 7 rebrand, is the only one still under meaningful development; it runs on reasonably fast Itanium hardware.

Is the mainframe dying? Maybe in a sense. NEC and Hitachi are probably going to pull out of the mainframe business in the next few years, perhaps followed by Fujitsu and Bull. Unisys and IBM will stick around for the simple reason that The Cloud makes the need for a huge, fast, hyper-reliable data-processing box more relevant, not less. IBM and Unisys have both had great sales numbers over the last couple of years, and I don't think that trend will stop any time soon. As long as mainframes stay fast and good at what they do, they can survive and thrive.