> Doesn't using kernel threads imply lots of context switches?

No, it's actually fewer context switches. That's because each I/O completion event goes straight from the kernel to the code that was waiting on it (1 context switch), not from the kernel to the userland dispatcher to the code that was waiting on it (2 context switches).

This is not quite right. The userland dispatcher to run next thread is not a context switch: the TLB doesn't need to get flushed for instance. Furthermore many userland threads can be woken at once, and don't incur context switches when they suspend, prompting the next runnable thread to run.

If you're willing to put in a lot of compiler work userland thread switching is a function call, literally. I don't 1:1 threads with small stack are going to compete with that.

The real cost of a true "context switch" is the transition from user level to kernel level, which takes thousands of cycles. This cost isn't incurred on a userland context switch, so those costs aren't comparable.

Thousands of cycles for a SYSCALL/SYSRET on a reasonably modern Intel/AMD CPU? I think you should try to measure that.

"Thousands of cycles" is definitely not true on modern CPUs. SYSENTER/SYSEXIT takes 150 cycles on my Xeon X3460.

Given the kind of "cloud"-backed startups most of HN are working on, I think the more practical measurement would be the overhead of a ring 0/3 separation in a VM, plus hypercalls, plus a ring 0/3 separation in the hypervisor (for paravirtualized syscalls); averaged against a ring 0/3 separation in a VM, plus SR-IOV virtualization (for HVM-backed syscalls.)

Yeah, this doesn't apply for context-switches that happen because of plain old pre-emption, but it does happen if the context-switch is because e.g. a network packet wants to arrive at another process in your VM than the one that's currently running.

I would imagine that there's a reason bare-metal IaaS providers have a business model. :)

My knowledge of current timings are admittedly a little out of date since my microkernel days are behind me. Even the 150 cycles another poster suggested is still a significant cost over a userland context switch.

What if you have async system calls (e.g. FlexSC)?