The Open Mainframe Project

Mainframes take a substantially different approach then the typical Unix-style 'big iron' that people are used to when they think of big databases and such things.

Mainframes are designed to be 'component' machines. You have the central CPU box which houses the main processors and disks to boot the system. Different machines are different, but the ones I used had a separate PC-based backplane that booted OS/2 Warp on Intel processors to provide console access to the S/390 machine that constituted the main part of the system.

You can call that a 'CPU Box' because that is all it really is there for. Just running the CPU and main memory.

All other components are running over a sort of 'network'. The network is based on packets that are based on the IBM punchcards from the 1960's. All communication happens over this sort of 'electronic punchcard' network.

As a example:

The disk units are housed in DASD units. These are big machines whose job it is is just to manage the 'file system' (although it would not be recognizable to a Unix user) and disk/data management stuff. It's 'firmware' is very sophisticated and it has lots of processing power in the DASD unit. This is very different then typical commodity hardware that tends to rely on the main OS and CPU for most of it's workload. It's sort of like SAN, but it's not independent from the main system.

All the mainframe components are like that. They are 'smart' with lots of processing power and connect to the cpu box via a proprietary network of sorts.

Because of this Mainframes have huge amounts of I/O processing power. They can move data at rapidly in massive quantities.

However their processing and memory capabilities are surprisingly low. This is because all the licensing costs (which are massive) are based on MIPS. The more mips you have the more thousands of dollars per month you have to pay for all the software and support you need. So people want to buy as little processing power as they can get away with.

This leads to a fundamentally different approach to software versus typical Unix-style (which Microsoft Windows counts as Unix-style) systems.

With commodity-based servers you now have massive amounts of CPU and memory capacity, but I/O sucks badly. Most people's servers have bottle necks through 1GbE or even 10GbE to get to data. Even if you have local storage it's still relatively low capacity. It can be very quick if you are using SSDs, but it's still relatively small quantities. So people tend to burn away CPU and memory to make programming easier/quicker and/or to compensate for the lack of I/O. Try to store everything in databases, use compression, design the software to generate information 'close to home' and such things. They don't think anything about 'well this is running slow, so lets get 4 cores instead of 2'.

With Mainframes you have very scalable I/O, but you want to keep the processing power as low as possible. You want to keep your system at 100% cpu load all the time to make the most efficient use of computer time and available resources. Your software is going to be much highly optimized. People have spent 30 or more years hand optimizing assembly code for processing stuff like addresses and purchase histories.

That sort of thing.

It's a pretty alien world. Very little is familiar. To a life long Linux user it makes Windows seem almost 'native' compared to the operating environment people typically use for mainframes.

Also because the software is so highly optimized backwards compatibility is cherished. I have used 1980's terminals on 2000-era mainframes. With tape drives from the late 1990's. We even had vertical tape drives that were from the 1970's that pre-dated sophisticated printed boards and instead used lacquered coated wire and hand-wound pin boards for their systems. All these things worked together seemlessly.

IBM is going into desperation mode really. Mainframes are their bread and butter and IBM is no longer the dominate '2000 pound guerilla' it once was. They have kept their stock price decent through weird and complex accounting schemes involving massive stock buy-backs. But they are starting to reach the end of that and the possibility of a real crash is on the horizon.

People have learned to work around commodity hardware I/O limitations through designing their software to be highly parallel across cluster machines (ie: Hadoop, etc). There are still workloads that can justify mainframe-style I/O capabilities, but they are few and far between. It is becoming increasingly difficult for businesses to justify keeping their mainframes going while I expect new sales of mainframe systems are almost now non-existent. If IBM's loyal customer base for their mainframes smell blood in the water then the mainframe is going to stop being a 'safe bet' and they will start to run away in droves and bite the bullet and dive into the world of java-based enterprise software on Windows or Linux systems.

Linux on mainframes can reduce costs, but Linux isn't really intended for high optimization that mainframes want to see. Even just having TCP/IP stack in software represents a massive loss in CPU capacity from their perspective. Anybody that buys into IBM's idea that mainframes are a suitable for running linux 'big data' workloads they are totally incompetent at their jobs. Running something as inefficient as 'zookeeper' or 'hdfs' on a mainframe is insane.