Re: [Ksummit-discuss] [TOPIC] Application performance: regressions, controlling preemption

[Andy Lutomirski <luto@amacapital.net>]

On Mon, May 12, 2014 at 7:32 AM, Chris Mason <clm@fb.com> wrote:
> Hi everyone,
>
> We're in the middle of upgrading the tiers here from older kernels (2.6.38,
> 3.2) into 3.10 and higher.
>
> I've been doing this upgrade game for a number of years now, with different
> business cards taped to my forehead and with different target workloads.
>
> The result is always the same...if I'm really lucky the system isn't slower,
> but usually I'm left with a steaming pile of 10-30% regressions.
>
> The KS dates should put us right at the end of our regression hunt, I can
> talk through the main problems we hit, how (if) we fixed them and hopefully
> offer tests to keep them from coming back.
>
> We've split the tiers up between Josef Bacik, Jens Axboe, myself and a few
> others.  I'd nominate Josef and Jens to come share their findings as well.
>
> Another topic here is controlling preemption from userland without needing
> to go full real time.  CPU intensive in-memory databases are leaving a ton
> of performance on the floor in context switches.  I'm hoping to experiment
> with better preemption controls and the userland RCU project to improve
> things.

I wonder how much of this overhead is the actual interrupt + context
switch and how much is overhead due to caching effects.  If the
former, there may be value in having a simple benchmark for how long
preemption takes and running it periodically.

--Andy