Re: Spinlock backoff algorithm

Greg,

> That says precisely nothing about the matter at hand. Someone should
> simply change it and benchmark it in pgsql. I doubt you'll see a
> difference there on regular AMD/Intel ... and if it makes the sun
> hyperthreaded cpu happier...

Nah, if it's only Niagara, it's not worth bothering. I was under the
impression that most x86 chips had similar issues around having less FP than
IP, but I could be wrong. It's also worth thinking about the new Intel
multi-core archtectures; do they starve FPs as well?

On a busy oltp system, spinlock waits get called 100's of times per second, so
like procarraylock this it's frequent enought to call for microoptimization.

I think maybe we should try making the change and testing it on Niagara and on
some standard x86 platforms, and then on the new x86 architectures, to see if
it makes any difference in CPU utilization.

FYI, if nobody had guessed this is coming out of study Magne is doing on
improving PostgreSQL SMP scalability.

--
Josh Berkus
PostgreSQL @ Sun
San Francisco