[lustre-discuss] Constant small writes on ZFS backend even while idle

[Alex Vodeyko]

Laura, Andreas, thanks a lot.

I've tested "local zfs" with multihost=off, so it's MMP.

Thanks again,
Alex

вт, 7 окт. 2025 г. в 20:03, Andreas Dilger <adilger at ddn.com>:
>
> It could be Multi-Mount Protection (MMP)?
>
> Cheers, Andreas
>
> > On Oct 7, 2025, at 08:52, Alex Vodeyko via lustre-discuss <lustre-discuss at lists.lustre.org> wrote:
> >
> > Hi,
> >
> > I'm in the process of testing lustre-2.15.7_next on rocky-9.6, kernel
> > 5.14.0-570.17.1.el9_6.x86_64, zfs-2.3.4.
> > 84 disk shelf, multipath.
> > 2x OSTs per OSS
> > OST is on the zpool with 4x(8+2)raidz2=40 hdds configuration (btw -
> > also tested on draid - the same problem).
> > atime=off (also tested with relatime=on)
> > recordsize=1M, compression=off
> >
> > During benchmarks I've found that even on the completely idle system,
> > zpool-iostat shows 40-160 4k (ashift=12) writes (1-4 per hdd) every
> > second.
> > # zpool iostat 1
> >              capacity     operations     bandwidth
> > pool        alloc   free   read  write   read  write
> > ----------  -----  -----  -----  -----  -----  -----
> > ..
> > ost00       482G   145T      0    158      0   634K
> > ost01       401G   145T      0     40      0   161K
> > ----------  -----  -----  -----  -----  -----  -----
> > ost00       482G   145T      0     40      0   161K
> > ost01       401G   145T      0    157      0   629K
> > ----------  -----  -----  -----  -----  -----  -----
> > ost00       482G   145T      0     40      0   161K
> > ost01       401G   145T      0     40      0   161K
> > ----------  -----  -----  -----  -----  -----  -----
> > ost00       482G   145T      0     38      0   153K
> > ost01       401G   145T      0     39      0   157K
> >
> > Could you please advise if I can turn off something (probably in
> > lustre, because local zfs does not show this behaviour) to avoid these
> > writes because they affect read performance (and cause huge cpu load
> > average and iowait numbers especially during multiple concurrent reads
> > from the single OST).
> >
> > Many thanks,
> > Alex
> > _______________________________________________
> > lustre-discuss mailing list
> > lustre-discuss at lists.lustre.org
> > http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org