Re: [tip:x86/mm] [x86/mm/tlb] 209954cbc7: will-it-scale.per_thread_ops 13.2% regression

[Rik van Riel <riel@surriel.com>]

On Mon, 2024-12-02 at 11:30 -0500, Mathieu Desnoyers wrote:
> 
> Or we just build a per-cpu mm_cpumask from per-CPU state
> every time we want to use the mm_cpumask. But AFAIU this
> is going to be a tradeoff between:
> 
> - Overhead of context switch at scale
> 
>    vs
> 
> - Overhead of TLB flush
> 
> 
> So I guess what we end up doing really depends which scenario we
> consider
> most frequent.
> 
I think that is going to be more workload dependent than
anything else.

If you're doing a kernel compile, or running a bunch of
shell scripts and simple Unix commands, you are dealing
mostly with single threaded programs, where not sending
IPIs is the best thing to do.

If you're running a long-lived, heavily multithreaded
program, you will benefit from reducing the context
switch overhead more than anything else.

Both seem like equally valid use cases.

I'm playing around with a patch now that builds on
my previous patches, but only trims the mm_cpumask
once a second.

Hopefully that can give us a reasonable medium between
the two.

-- 
All Rights Reversed.