On Mon, Jul 8, 2019 at 2:42 PM Maged Mokhtar <mmokh...@petasan.org> wrote:
> > On 08/07/2019 13:02, Lars Marowsky-Bree wrote: > > On 2019-07-08T12:25:30, Dan van der Ster <d...@vanderster.com> wrote: > > > >> Is there a specific bench result you're concerned about? > > We're seeing ~5800 IOPS, ~23 MiB/s on 4 KiB IO (stripe_width 8192) on a > > pool that could do 3 GiB/s with 4M blocksize. So, yeah, well, that is > > rather harsh, even for EC. > > > >> I would think that small write perf could be kept reasonable thanks to > >> bluestore's deferred writes. > > I believe we're being hit by the EC read-modify-write cycle on > > overwrites. > > > >> FWIW, our bench results (all flash cluster) didn't show a massive > >> performance difference between 3 replica and 4+2 EC. > > I'm guessing that this was not 4 KiB but a more reasonable blocksize > > that was a multiple of stripe_width? > > > > > > Regards, > > Lars > > Hi Lars, > > Maybe not related, but we find with rbd, random 4k write iops start very > low at first for a new image and then increase over time as we write. If > we thick provision the image it work does not show this. This happens on > random small block and not sequential or large. Probably related to > initial obkect/chunk creation. > object_map can be a bottleneck for the first write in fresh images -- Paul Emmerich Looking for help with your Ceph cluster? Contact us at https://croit.io croit GmbH Freseniusstr. 31h 81247 München www.croit.io Tel: +49 89 1896585 90 > > Also we use the default stripe width, maybe you try a pool with default > width and see if it is a factor. > > > /Maged > > _______________________________________________ > ceph-users mailing list > ceph-users@lists.ceph.com > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com >
_______________________________________________ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com