ADVENT #23 — bernoulli
I really wanted to write about the Canon Optical Card, but I’ve never even seen one. Probably Popular Science wrote about it as the data storage of the future just before, or just after Canon abandoned it. I probably read about it in a tattered copy a decade after that. I liked it for this series because it was the kind of write-once media that would be perfect for Alice if it were as cheap as dirt. It never was cheap though. Could it have been cheap? Maybe. What makes things cheap isn’t wanting them to be cheap, it’s competing with somebody who thinks they can be made cheaper. I guess that’s half of capitalism. What makes things expensive? Figure that out and you‘ve got the other half.
I was maybe instead going to write about holographic storage but after a career of waiting for that to be a thing, it’s as much the storage of the future as Dippin’ Dots are the ice cream of the future. Heck, holographic storage may still be the next big thing years after we start using clean nuclear fusion to make ice cream.
The Bernoulli Box is a technology that I do have personal experience with. In the early days of hard drives, and the middle days, and a bit into the later days hard drives were bedeviled by head crashes. With tapes and floppies, the magnetic head slides along the media and can eventually wear through the recording layer. Hard drive heads fly above their media right until they don’t. For all our fascination with transistors that are just a few nanometers on a side, hard drives heads been flying within a transistor’s width of the media since they were first built. Drives can’t spin forever though and that means the heads need to eventually land somewhere. When they land someplace safe, we say they are parked. Otherwise, they maybe took out some of your data. If head kicks up some material that circulates, then the wreckage goes on. Starting about 40 years ago, hard drives parked themselves when power failed. Starting almost 25 years ago, accelerometers in laptops like Apple’s Titanium PowerBook G4 automatically directed hard drives to park themselves if the laptop detected a fall.
A startup called Iomega saw a way through this with a floppy-like media media that flew above the heads and drew them close to the disk with a low-pressure region induced between the head and the media. Bernoulli’s principle gave this new drive an intrinsic safety mechanism and a handy name. The Iomega Bernoulli box had few downsides. It really did have almost floppy like convenience with hard drive like capacity. They were especially useful as off-site backup cartridges for valuable business records.The problem is … that the problem the Bernoulli box solved just sort of solved itself. Hard drives became more reliable in the way that volume makes this more reliable, and larger enough that Bernoulli boxes were uncompetitive. Iomega wasn’t yet finished — it would take several more swings at the market with Zip (based on floptical technology profiled earlier), Jaz, and Click! drives.
There’s not much connection to Alice with Bernoulli except that a backup medium loses all utility when its rate of technological progress falls behind the primary storage media. Since backups are valued less and by fewer people than primary storage, this happens again and again. There is no stable point except for them to be the same technology, which means backups have to be continually propagated onto newer media. Alice does the same thing and manages locality on a global tape by copying useful material forwards from time to time. My furnace code does the same thing. Important state is brought forward in the journal so that about 1% of the journal is just material brought forward from earlier spots in the tape as it‘s used. It makes it easier to restart the system without replaying from the beginning, but it also makes sure that data is incrementally backed up along the way.
The perfect storage system for Alice is one that takes advantage of some new economy, not necessarily some new material science. The only way to bootstrap a write-once ecosystem is for it to devour media so quickly that the media has no choice but to be cheap. The media probably has to serve some other function in the system as well — as fuel or a heat transfer media. One of the things that keeps cutting tools cool is that they eject hot chips that carry away some of the heat of the operation. Perhaps some future system can use a streaming media the same way.
I’m still working to finish the tracer for Alice that’s supposed to be part of this system. I guess the problem is that I haven’t needed it. I just debug my Alice programs the same way I have been debugging them for 40 years, with something like print. The way to get a good trace is to make the system need it for its own operation. The best way to do that is to checkpoint less information into the tape when threads block and instead to recover that information from the tape. If I can do that, then I can increase the number of general-purpose registers without increasing the amount of thread state written to the tape, and likewise reduce the amount of thrashing on the tape just to manage the juggling of only two registers.
By the way, the IBM JX from the last installment was neither IBM’s first attempt nor their last attempt to converge the Japanese and US PC markets. The original IBM PC and clones struggled against the homegrown PC-98 standard in Japan. IBM tried again a couple of years later with the Japan market PS/55 system that was substantially the same as contemporary PS/2 systems sold in the US. 1999 brought the first portable PS/2 (US: PS/2 P70, Japan: PS/55 5545-T) achieved convergence in everything but keyboard layout — and meant also that IBM had finally mastered production in Japan of a machine for the US market. This success led to the PS/55 Note series, a Japan-first series of notebook computers that became the basis for the ThinkPad line.