Is this any better than the Docker method of reusing the same base-OS and compartmentalizing the applications? Is there that much to be gained in avoiding kernel/user-space transitions?
If you want to bring the isolation level of that process down to just absolutely what it needs to run we've got things like jails and cgroups. You could probably run a Go app with no access to the filesystem since everything is linked in statically anyways.
I think it misses the reasons people are excited about virtualization. Reproducibility and uniformity of environment has a higher value than isolation to most software developers. The priorities may be inverted on the sysadmin side, but I don't think so far as to justify this kind of approach.
You can achieve the isolation with jails and cgroups, but not the performance improvements.
And this is what I mean when I say that taken to its conclusion you're just reinventing processes.
I think this kind of performance claim needs to be solidly proven by something at least vaguely like a real running application to be taken as a given.
However, as Docker PaaS gains popularity, that may be a better alternative. Only benchmarks will tell :)
Shouldn't the unikernel approach actually improve reproducibility greatly? You build your application and all it's dependencies together, that should run exactly the same locally or on your Xen cloud.
An example of this is the Mirage website itself, where all the kernel outputs that are live are stored in GitHub at https://github.com/mirage/mirage-www-deployment -- an explanation of the Travis CI workflow is at http://www.openmirage.org/wiki/deploying-via-ci
Yes, there's at least an order-of-magnitude improvement in packet processing, for example, if you bypass the kernel.
Our mobile app runs WebSocket-like connections over UDP with libsodium for crypto, and we're moving our stack off of Node.js for exactly that reason.
Things like file descriptor limits or numbers of connections are more of a pain in modern times, not necessarily a context-switch caused problem.
That thing they were patching? That's the "data plane". It was incredibly high bandwidth relative to the control plane, since it was completely optimized for moving data.
Back to Unix. Unix is designed for the control plane. It is not designed for rapidly moving large amounts of data with maximum efficiency. Why do we use it today for what are arguably "data plane" tasks? Well, when all you have is a hammer...
Nowadays, you can have both on the same machine: run Linux on the first core or two, and reserve the remaining cores for your app. Along with huge page allocations to reduce TLB impact, you can literally own all CPU activity on those cores, and lock all of your RAM too.
That app is a normal Linux app, but it runs on the raw hardware—like not having an operating system at all. When you also give your app complete control over the network hardware, you've completely bypassed the kernel.
With UDP, you don't even need a networking stack, making this approach particularly attractive. Another poster mentioned saturating a 10Gb link. How about saturating four 10Gb links on a single machine? It can be done with the E5 processors and the software architecture I described above.
We're shooting for 10 million packets processed per second on a single, ~$20K machine. That's pretty sweet if you ask me, and a hell of a lot more than Node.js can do.
Once that's done, the Xen backend is just a matter of filling in the missing kernel components with OCaml libraries (or, in the case of OSv, with C libraries, or in HalVM's, Haskell libraries).