Re: [PATCH v3] sched/fair: Add advisory flag for borrowing a timeslice

[Davidlohr Bueso <dave-h16yJtLeMjHk1uMJSBkQmQ@public.gmane.org>]

On Tue, 2014-11-25 at 12:23 -0700, Khalid Aziz wrote:
> On 11/25/2014 11:27 AM, Davidlohr Bueso wrote:
> > On Tue, 2014-11-25 at 07:45 -0700, Khalid Aziz wrote:
> >> This solution has been used by both database and java on other OSs and
> >> has shown performance improvement. Andrew had asked for performance
> >> numbers on Linux with this patch last time I sent this out and it took
> >> me a while to get performance folks to run a full TPC-C workload. They
> >> did see a 3% improvement in tpcc as I noted in commit log and that is a
> >> significant improvement.
> >
> > 3% for such a change seems pretty worthless... I would have expected
> > this having to impact performance much more.
> >
> 
> Performance impact will depend upon how big a bottleneck the spinlock 
> was creating and how severe the resulting convoy problem was. Database 
> guys try to squeeze every bit out of the system and 3% is considered to 
> be very good gain. 10% gain would have been nicer :)

Right, but my point is that 3% indicates that this isn't really a
problem in the first place. It would be good to know that hw
characteristics as well. We've tackled Oracle related performance issues
in the past with on OLTP benchmarks, with much more noticeable
improvements -- which is why I'm not impressed with your numbers and
particularly for a patch of this nature. That said, I do realize that
Oracle is not the only workload that can potentially gain with userspace
spinlocks.

Thanks,
Davidlohr