You need an event-driven system because the number of "tasks" rarely matches the number of physical cores available. One tried and true, fault-tolerant event-driven interface is provided by the kernel: processes (and threads). It has interfaces for binding to NUMA regions and such.

Any system you design will have to tackle a lot of the hard problems that kernels deal with. Now maybe you have a simpler system in mind, but it has to be drastically better for anyone to consider giving up binary compatibility.

You are completely misunderstanding me. I am not saying "let's all move to this new utopian system". I am suggesting that we move away from the idea of making universal systems entirely. Write software that is small enough that complete rewrites can be accomplished without trouble. The current situation is to build ourselves a problem that gets harder and harder to undo as time goes on.

You can't call something that's been around for less than 100 years "tried and true". I'm not saying we can't have "virtual CPUs" for tasks that can handle a performance hit. But it is becoming an unavoidable impediment. "Simple" doesn't mean "easy". It means predictable, small and undoable. The exact opposite to where we are heading.

I write a lot of code that runs in funny environments like the Compute Node Kernel running on Blue Gene systems, with offset-mapped memory (no TLB) and no over-subscription. It's fantastic for reproducible performance, but it's a specialized environment and people want to do things like run Python (as "glue" for some scientific applications) which needs dynamic loading. Since all IO is exported to separate IO nodes, and due to file system consistency semantics, dynamic loading is a tremendous bottleneck. This has escalated to the point where people burn a quarter million core hours to load Python (plus C extension modules) once on a large machine. To really solve the scalability problem, we needed to make dlopen() avoid touching the file system (by patching ld.so and implementing the POSIX file API with collective semantics served over the fast network).

The point here is that there are good reasons to want to use different systems together and then you end up putting in lot of effort working around the limitations of the specialized environments.

It sounds like you're basically proposing Plan 9--lots of very small, single-purpose components with one standardized way of passing data between them used across the entire system.

Also, you can ask a lot of people to give up binary compatibility and be okay, but asking people to give up source compatibility and change applications is a much, much harder pill to swallow. Specialized environments (like the syscall-shipping CNK on Blue Gene like jedbrown mentions) are usually seen as something to work around because they are impediments to productivity. A more robust system wouldn't be considered useful if everyone has to start from scratch.

I'm not really interested in what's "considered useful". Look at the state of web applications. Basically what we could already do in 1996 but using more resources and full of ads and other distractions. And yet to many people this step backwards represents state of the art computing. "Productivity" is highly overrated. Computers are already extremely useful without a bunch of gimmicks. Most of the Web 2.0 companies are creating the desire for their own product. Facebook has turned family photo sharing into a circus show. It doesn't serve any need other than to push for more neomaniacal computerisation. Some people will want to share family photos over the web. We can do that with very simple software. It can even be loaded on demand and sandboxed without any configuration without a hugely complex platform. _Simple_ virtual machines capable of this kind of work aren't hard to construct. But people keep chasing idiotic benchmarks instead. And if it's too slow then change the hardware to make sandboxing easier.

How do you handle DMA and other I/O in your system?

Each hardware device is accessed by a single process at a time. Any multiplexing that needs to be fast is done in hardware. Other multiplexing can be done in user space. For example, and application can request a dedicated channel to some blocks on the hard disk (adjudicated by the OS) or just use another hard disk. Remember, we're building simple applications here so things don't need to constantly write off the the disk for various tasks. You would only be dealing with the disk if your actual data was too big to fit into memory. Cases like that need careful design, not delegation off to some other programmer who doesn't know your problem.

If you're running a server application you split it into some reasonable number of tasks and allocate cores to them as needed. Nothing else runs on the system at all. If you need more cores either upgrade your hardware or you build another identical machine and talk to it over the network. If you look at the way Moore designs, he's crafting a complete artifact to solve a specific problem, not just throwing generic parts together.