Re: [PATCH] x86/asm/entry/64: pack interrupt dispatch table tighter

[Ingo Molnar <mingo@kernel.org>]

* H. Peter Anvin <hpa@zytor.com> wrote:

> On 04/03/2015 11:37 AM, Linus Torvalds wrote:
> > On Fri, Apr 3, 2015 at 11:35 AM, H. Peter Anvin <hpa@zytor.com> wrote:
> >>
> >> For the record, I actually measured the impact of the jump-to-jump when
> >> I wrote it.  It has a small, *but measurable*, positive impact.
> > 
> > What did you compare against, and how did you measure that? I don't
> > see how it could *possibly* be faster than just a simple aligned "push
> > + jmp".
> > 
> 
> I wish I remembered the exact details; it took a fair bit of gathering
> numbers as the spread was quite a bit wider than the delta, but in the
> end there were two distribution peaks clearly offset.
> 
> I seem to remember it involving a loop running RDTSC continuously and
> another RDTSC in the interrupt path.

So the thing is, while I don't know how you've loaded the machine, if 
user-space is doing nothing but looping in RDTSC, the kernel I$ might 
become cache hot easily not just in L2 but in L1 cache as well.

But 'when the machine is not doing anything' is not what we optimize 
for, we (try to) optimize for the case when there's a lot of work 
going on and the async context (irq, fault, etc.) I$ is likely 
cache-cold.

Thanks,

	Ingo