Re: [PATCH] NUMA spinlock [BZ #23962]

[Torvald Riegel <triegel@redhat.com>]

On Tue, 2019-01-15 at 13:17 +0100, Florian Weimer wrote:
> * Torvald Riegel:
> 
> > What I mean is that applications would have to want to use locks provided
> > by tbb, whether those are locks/mutexes that exist in tbb today or a new
> > API that would be added.
> > 
> > Put differently, I'm not optimistic about tbb being a good way to get
> > feedback.
> 
> Do you want to run existing workloads with a new mutex implementation?

We need to get there.

> Then we can't add new flags or change ABI in any way

Yes.

> and would have to
> use a tunable.  And to get feedback, we would have to make the new
> implementation the default, with a tunable to get back the old
> implementation.

I wouldn't be too concerned to getting back the old implementation, so
maybe we don't even need a tunable right now.  The old implementation is
just no spinning, so the cases where I can imagine the tunable to be useful
is either (1) experimentation to compare performance without using
different glibc's and (2) going back to old behavior in cases where we
really screwed up.  But how many users will have the time to investigate
(2), really?

Nobody should have to tune their spinlocks, or the back-off in mutexes. 
It's our duty to make sure this has good average-case performance.