Re: CFQ timer precision

[Jens Axboe <axboe@kernel.dk>]

On 11/16/2015 08:11 AM, Jan Kara wrote:
> Hello,
>
> lately I was looking into a big performance hit we take when blkio
> controller is enabled and jbd2 thread ends up in a different cgroup than
> user process. E.g. dbench4 throughput drops from ~140 MB/s to ~20 MB/s.
> However artificial dbench4 is, this kind of drop will likely be clearly
> visible in real life workloads as well. With unified cgroup hierarchy
> the above cgroup split between jbd2 and user processes is unavoidable
> once you enable blkio controller so IMO we should accomodate that better.
>
> I have couple of CFQ idling improvements / fixes which I'll post later this
> week once I'll complete some round of benchmarking. They improve the
> throughput to ~40 MB/s which helps but clearly there's still a big room for
> improvement. The reason for the performance drop is essentially in idling
> we do to avoid starvation of CFQ queues. Now when idling in this context,
> current default of 8 ms idle window is far to large - we start the timer
> after the final request is completed and thus we effectively give the
> process 8 ms of CPU time to submit the next IO request. Which I think is
> usually far too much. The problem is that more fine grained idling is
> actually problematic because e.g. SUSE distro kernels have HZ=250 and thus
> 1 jiffy is 4 ms. Hence my proposal: Do you think it would be OK to convert
> CFQ to use highres timers and do all the accounting in microseconds?
> Then we could tune the idle time to be say 1ms or even autotune it based on
> process' think time both of which I expect would get us much closer to
> original throughput (4 ms idle window gets us to ~70 MB/s with my patches,
> disabling idling gets us to original throughput as expected).

Converting to a non-jiffies timer base should be quite fine. We didn't 
have hrtimers when CFQ was written :-)

-- 
Jens Axboe