I can only read the abstract. It mentions that applications need to be recompiled. Would that mean applications when in dotnet or java wouldn’t need modification, just their runtime would?

The full paper is here: https://dl.acm.org/doi/pdf/10.1145/3552326.3587458

Yes, everything has to be recompiled, because it must be linked to the Linux kernel to create a single vmlinuz containing application + kernel.

One of the aims[1] was requiring minimum source-level modifications to achieve this. For C/C++ programs this was pretty much achieved. As you say for Java or Dotnet you'd have to port the runtime (which we never did) and then the class files ought to run unmodified on top.

Having said that, there's not much benefit to this (perhaps a negligible increase in performance) if you only link your application to Linux. The real benefits come when hotspots within your application are modified so they interact directly with Linux internals (noting that Linux internal APIs are not stable). One example from the paper is shortcutting through the VFS when you know that a read(2) call is always reading from a TCP socket. Another might be to arbitrarily modify page tables to make your garbage collector more efficient. Another one which my student was working on was making Linux epoll handlers direct call into the "userspace" code that handles an event.

A cute part of this is that it's still Linux, just with a "fatter" kernel containing the application. It still has a userspace, and you can still run stuff in that userspace if you want. That's especially useful for debugging (put sshd + a shell in there), or performance analysis (run perf alongside the unikernel program).

There are definitely arguments to be had about whether any of this is a good idea or not, which is why it's a research paper. I think nowadays I'd say that much of the performance benefit can be achieved using io_uring.

[1] Another aim was making minimum changes to Linux itself so it might have some chance to go upstream, which of course it never did.