On Thu, 2018-05-31 at 14:59 -0400, J. Bruce Fields wrote:
> On Wed, May 30, 2018 at 02:05:34PM -0400, Trond Myklebust wrote:
> > This patchset is something we've been sitting on for a couple of
> > years
> > in order to allow it to mature. It was mainly developed by Fred
> > Isaman
> > while he was working for Primary Data, but we've had a few bugfixes
> > come in since then.
> > The main reason for delaying the upstream merge was to ensure that
> > adding layoutget to the OPEN compound was safe (does not cause
> > existing
> > layouts to be automatically revoked) and that it gives us a real
> > performance benefit.
>
> Do you have any of the performance data?
>
> Or just a rough summary explaining when it helps most and by how
> much?
>
It's very workload dependent:
If you have large files that are being written from scratch, then the
layoutget on OPEN makes no difference, since the overhead of a single
layoutget to the metadata server is quickly amortized.
The main difference is with workloads involving lots of small files,
where you want to open, then send one or two writes, then close; if you
can eliminate one round trip to the MDS, then that can improve
performance by a bit (not a huge amount). Ditto for workloads where you
open, then send a read, then close, however read caching can quickly
eliminate any advantage of sending a pre-emptive layoutget here.
--
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@xxxxxxxxxxxxxxx
��.n��������+%������w��{.n�����{��w��� jg��� �����ݢj����G���� ���j:+v���w�m������w�������h�����٥
