From: Jeff Dike <jdike@ka...> - 2002-02-12 20:48:13
> The 50% "loss"; is that real loss or performance just not available
> inside UML but "idle" on the host (which might mean that 2 UMLs would
> be able to deliver ~100% of the host performance total)? I recall that
> the 50% "loss" was mostly due to scheduling overhead.
It's both. If you ^C UML during a boot (or any other I/O intensive load),
you will see that you are frequently in the idle loop. This is because of
the block driver being limited to one outstanding request at a time, which
badly hurts readahead attempts. Adding AIO support will help this.
The rest is lost (in the sense that current host kernels can't do any better
for UML). I have a bunch of schemes in mind to fix the host that should
grab back a lot of that loss.
> Second, might Ingo's latest scheduler improve that number?
I don't think so, but feel free to try and report results. I've been
surprised before (by the ~2X speedup of tmpfs /tmp vs ext2 /tmp).