Late-night braindump!

Guys! We've totally missed the real potential with ChromeOS!

It's got potential as a netbook OS, sure. But, beyond that, appliance computing - embedding the OS in the hardware is a huge deal! Especially if it's stateless. Put two copies of the OS on separate flash drives, glue them in, reduce the moving parts (how many do you really need, anyway? OLPC got it down to 0, and something like 3 internal plugs), and you've got a 100% self-contained system that never needs maintenance! As long as we're borrowing tricks from OLPC, you could put a battery on it and have it portable too! Add a hand-crank and you've got truly mobile computing. Everything else along this line of thinking is just implementation detail, but yeah!

An interface (monitor, input devices, whatev) would be nice, but even nicer would be a system that's implicitly based on clustering. You'd take two or three or four of these, plug them into a switch, plug a monitor and a hard drive or two into that same switch, and you've got a computer that's resistant against hardware faults and completely network-based. You could upgrade it just by plugging new components in and... that would be it, yeah! Probably wouldn't even need a restart. We already have the foundations for zero-configuration networking, so this would be the easy part.

Oh hey! as long as it's network based, you could add multiple user terminals and have several people using a shared cluster with full access to all resources. I think most of the problems here are solved; the interesting one I see is networked storage - you'd need encryption, because we have to assume that anybody can see any hard drive, and hmm. That might actually be it. Obviously you'd need a way to provide a private key - maybe adding a USB port to the input devices and plugging in a flash drive. Idea! The input device could be a thin client with a custom interface, which you could carry around and plug in to any cluster. That would take a lot of the burden off the OS on the cluster nodes (which would have to be pretty stripped-down and generic to begin with) and oooh. That'd be neat.

Problem: untrusted clusters. Crypto-computation might be impossible (I have seen some laughably bad research attempting to solve it), so we might have to assume that any given node can potentially snoop on computations and alter them. It might be possible to finagle something with a signed system on the cluster nodes, but they'd be too much of a high-value target to trust that some other kind of manipulation wasn't going on. Trust is a fundamentally hard problem, let's ignore it for now and work around it. Either we implicitly trust the cluster (boring, stupid) or we figure out a way to run computations on untrusted nodes so that - dammit, back at crypto-computation. :[

If Light Peak lives up to its potential, it would be kickass for this. Everything would only need one plug. Problem: fiber is a lot less resilient than CAT-5. Ah, well, not like the layer-1 matters too much.

Appliances require maintenance sometimes! What we really want here is a computer that you can embed in the wall and forget about, and have as a seamless part of the network doing useful stuff, and be reasonably confident that it'll outlast the building. I can't think of a name for this right now, but it'd be pretty cool. Transparent computing? Ubiquitous computing? Ooooh! As long as we're sticking these in the walls, we could create a wireless mesh network with the neighbors and if this actually becomes widespread it'll be usable for stuff. Routing in mesh networks is hard and really outside the scope of this braindump but the potential payoff is too cool for me to ignore.

Anyway, back to transparent networked storage! I am imagining an interface where you can just dump data onto a network, and "pull" all data owned by you back onto your own system when you leave the network. This would be at least a billion times nicer than our current home network storage metaphor, which is complete crap. Network storage shows up as hard drives, which sometimes vanish, and may or may not be accessible based on remote configurations, which can change at any time, and concurrent access is a bad joke, and seriously who came up with this ridiculousness? But I digress. We want all storage to be accessed in a unified way, and always writable would be nice - a reciprocal protocol would guard against people just hogging all the storage, but dunno how that'd work yet. Future braindump topic! Maybe require people to provide all the storage they actually use, but distribute the data around the network so they get the benefit of insurance against disk failure.

(For any of this to be feasible, we have to be able to build 100% secure operating systems. Attention, programmers everywhere! Stop sucking. :( )

If we could set this up in such a way that you could charge people for using resources on your nodes, we'd have a completely distributed clone of EC2. Cool! (Oh, hey, that'd work for storage too. Two birds with one distributed ripoff! :D)

Back to the software side: to share compute resources, we'd need a virtual machine and a stable OS API so that different computers can use the cluster. There are a few good choices right now (JVM and CLR mainly, and LLVM is looking good). Randomly: it'd be nice to have a VM that supported distributed execution of a single program. Might get messy with failures -- no, wait, the VM should handle failures! And the program should just see a single (highly parallel) machine that never fails! Gosh, that'd be neat.

Years of utterly worthless computer security have left us in a state where we can't even pass a file between two computers without bypassing six different "security features" designed to keep computers from ever communicating, just in case one side or the other is infected with a virus. This is pathetic! Hardware is improving at an incredible rate, but that hardly matters when software is completely and pathologically unable to keep up. The true cost isn't that some things are harder than they should be; sending files at all is a parlor trick. The true cost of poor security is all the cool stuff we could be doing if we didn't have to treat every damn network service as an attack vector.

Technology is pretty neat and all, but sometimes I wonder that we don't fully appreciate how primitive this all really is.