Re: track_planning causing performance regression

On 2020/06/30 20:30, Ants Aasma wrote:
> On Tue, 30 Jun 2020 at 08:43, Fujii Masao <masao(dot)fujii(at)oss(dot)nttdata(dot)com <mailto:masao(dot)fujii(at)oss(dot)nttdata(dot)com>> wrote:
>
> > The problem looks to be that spinlocks are terrible with overloaded CPU and a contended spinlock. A process holding the spinlock might easily get scheduled out leading to excessive spinning by everybody. I think a simple thing to try would be to replace the spinlock with LWLock.
>
> Yes. Attached is the POC patch that replaces per-counter spinlock with LWLock.
>
>
> Great. I think this is the one that should get considered for testing.
>
> > I did a prototype patch that replaces spinlocks with futexes, but was not able to find a workload where it mattered.
>
> I'm not familiar with futex, but could you tell me why you used futex instead
> of LWLock that we already have? Is futex portable?
>
>
> Futex is a Linux kernel call that allows to build a lock that has uncontended cases work fully in user space almost exactly like a spinlock, while falling back to syscalls that wait for wakeup in case of contention. It's not portable, but probably something similar could be implemented for other operating systems. I did not pursue this further because it became apparent that every performance critical spinlock had already been removed.
>
> To be clear, I am not advocating for this patch to get included. I just had the patch immediately available and it could have confirmed that using a better lock fixes things.

Understood. Thanks for the explanation!

Regards,

--
Fujii Masao
Advanced Computing Technology Center
Research and Development Headquarters
NTT DATA CORPORATION