Re: [PATCH 0 of 9] x86/smp function calls: convert x86 tlb flushes to use function calls [POST 2]

[Ingo Molnar <mingo@elte.hu>]

* Jeremy Fitzhardinge <jeremy@goop.org> wrote:

> This series:
>  - adds a simple debugfs profiling entry for cross-cpu tlb flushes
>  - converts them to using smp_call_function_mask
>  - unifies 32 and 64-bit tlb flushes
>  - converts smp_call_function to using multiple queues (using the now
>    freed vectors)
>  - allows config-time adjustment of the number of queues
>  - adds a kernel parameter to disable multi-queue in case it causes
>    problems
> 
> The main concern is whether using smp_call_function adds an 
> unacceptible performance hit to cross-cpu tlb flushes.  My limited 
> measurements show a ~35% regression in latency for a particular flush; 
> it would be interesting to try this on a wider range of hardware.  I 
> gather the effect tlb flush performance is very application specific 
> as well, but I'm not sure what benchmarks show what effects.
> 
> Trading off agains the latency of a given flush, the smp_function_call 
> mechanism allows multiple requests to be queued, and so may improve 
> throughput on a system-wide basis.
> 
> So, I'd like people to try this out and see what performance effects 
> it has.

nice stuff!

I suspect the extra cost might be worth it for two reasons: 1) we could 
optimize the cross-call implementation further 2) on systems where TLB 
flushes actually matter, the ability to overlap multiple TLB flushes to 
the same single CPU might improve workloads.

FYI, i've created a new -tip topic for your patches, tip/x86/tlbflush. 
It's based on tip/irq/sparseirq (there are a good deal of dependencies 
with that topic).

It would be nice to see some numbers on sufficiently SMP systems, using 
some mmap/munmap intense workload.

	Ingo