Jason J. Gullickson

Jason J. Gullickson

A Hardware Operating System?

Reviving the VROS project has radiated a number of new ideas, some relevant to the immediate work, some to what’s next and some... further out in the future.

As I’ve worked my way down to finding the most fundamental ways of putting pixels on the screen in Linux, I’ve been careful to avoid anything too specialized. One of the requirements of the work is to preserve portability, as my long-term plan is to produce a stand-alone operating system even if for now I’m using the Linux kernel as an interim solution.

Exploring a space with these constraints in mind while researching how to build a networked virtual environment yields a lot of problems to solve. Many of the existing solutions to these problems are compromises due to accepted limitations of tooling dependencies, operating system compatibility or simply the limits of the underlying hardware architecture itself. Since my goal is to produce a new operating system and ultimately a reference hardware design, I have a lot more latitude when it comes to solving these problems.

One example is updating the display in a timely fashion while servicing multiple input devices which move objects around in the virtual space. A pure software approach to this might block the display rendering while reading each input device in a round-robin fashion. A more sophisticated implementation might use threads and an event-driven model to allow each input device to update its position when the user moves. Either way at the hardware level the processor is still doing one thing at a time. Even in a multiprocessor system, there is usually a central thread that has to coordinate the work of each core or bus limitations that prevent completely parallel operation.

This lead me to imagining what implementing something like this in hardware might look like using an FPGA. The result was something akin to how a mainframe works, using separate processing units for each input/output “channel”, each with a dedicated connection to a memory where the virtual world “state” is stored.

It’s akin to implementing the threading and scheduling purely in hardware, which got me thinking: how much of the operating system I’m designing could be implemented in FPGA-programmed logic, and has something like a pure-hardware operating system been done before?

To answer that question we have to get specifics about exactly what an operating system is. You could argue that the ROM BASIC in the 8-bit micros from the early 80’s was an operating system, and at some level this was technically implemented in hardware (after all ROM chips are just arrays of programmable gates), but the traditional term for that is firmware. I think what I’m describing is distinct from a ROM-based O/S in that what I’m imagining is synchronous logic, not instruction code fed to a single CPU for execution.

(a diagram might be helpful here)

Of course implementing something on the scale of a Unix system in this way would be hard to imagine, but what I have in mind for VROS is orders of magnitude simpler than Unix, and it’s almost exclusively event-driven nature is actually simpler to implement in hardware than it is in software. Responding to hardware interrupts, controlling access to memory, performing mathematical calculations, these are all operations that have been delegated to specialized hardware over the years. In the case of VROS, this doesn’t leave a lot left to implement in the operating system software.

One remaining advantage to a software operating system is the ability to easily upgrade the system, even without physical access or new hardware. Traditionally this was a major advantage, but with an FPGA-based system, this sort of upgrade is possible even for hardware.

So far this idea is little more than a thought experiment, and the need for such a thing is at least three or more iterations down the road for what is needed for VROS. But I found the idea intriguing enough that I wanted to make sure I surfaced it somewhere so that if nothing else I could return to it again when time and resources allow.