Is there a way to make Cephfs kernel client to send data to Ceph cluster smoothly with buffer io

Is there a way to make Cephfs kernel client to send data to Ceph cluster smoothly with buffer io

[Sage Meng]

Hi All,

      Cephfs kernel client is influenced by kernel page cache when we
write data to it,  outgoing data will be huge when os starts to flush
page cache. So Is there a way to make Cephfs kernel client to send
data to ceph cluster smoothly when buffer io is used ? Better a way
that only influence Ceph IO not the whole system IO.

Re: Is there a way to make Cephfs kernel client to send data to Ceph cluster smoothly with buffer io

[Jeff Layton]

On Thu, 2020-11-12 at 23:17 +0800, Sage Meng wrote:
> Hi All,
> 
>       Cephfs kernel client is influenced by kernel page cache when we
> write data to it,  outgoing data will be huge when os starts to flush
> page cache. So Is there a way to make Cephfs kernel client to send
> data to ceph cluster smoothly when buffer io is used ? Better a way
> that only influence Ceph IO not the whole system IO.

Not really.

The ceph client just does what the VM subsys asks it to do. If the VM
says "write out these pages", then it'll do it -- otherwise they'll just
sit there dirty.

Usually you need to tune things like the dirty_ratio and
dirty_background_ratio to smooth this sort of thing out, but those are
system-wide knobs.

Another alternative is to strategically fsync or syncfs from time to
time, but that's sort of outside the scope of the kernel client.
-- 
Jeff Layton <jlayton@kernel.org>

Re: Is there a way to make Cephfs kernel client to send data to Ceph cluster smoothly with buffer io

[Sage Meng]

It would be nice if there was a mount option to have such parameters
adjusted that only influence CephFS Kernel Client.

Jeff Layton <jlayton@kernel.org> 于2020年11月12日周四 下午11:59写道：
>
> On Thu, 2020-11-12 at 23:17 +0800, Sage Meng wrote:
> > Hi All,
> >
> >       Cephfs kernel client is influenced by kernel page cache when we
> > write data to it,  outgoing data will be huge when os starts to flush
> > page cache. So Is there a way to make Cephfs kernel client to send
> > data to ceph cluster smoothly when buffer io is used ? Better a way
> > that only influence Ceph IO not the whole system IO.
>
> Not really.
>
> The ceph client just does what the VM subsys asks it to do. If the VM
> says "write out these pages", then it'll do it -- otherwise they'll just
> sit there dirty.
>
> Usually you need to tune things like the dirty_ratio and
> dirty_background_ratio to smooth this sort of thing out, but those are
> system-wide knobs.
>
> Another alternative is to strategically fsync or syncfs from time to
> time, but that's sort of outside the scope of the kernel client.
> --
> Jeff Layton <jlayton@kernel.org>
>