Slow nfsd write performance, tweaks needed

[John-Mark Gurney]

Mehmet Erol Sanliturk wrote this message on Thu, Nov 13, 2014 at 05:14 -0800:
> On Thu, Nov 13, 2014 at 5:00 AM, Rick Macklem <rmacklem at uoguelph.ca> wrote:
> 
> > Mark Schouten wrote:
> > > Hi,
> > >
> > >
> > > I am in the process of switching from a ZFS On Linux-based NFS-server
> > > to a FreeBSD-based NFS-server. The FreeBSD implementation of ZFS is
> > > way superiour over ZoL, and the box serves as storage for a
> > > virtualizationplatform, so stability is welcome. :)
> > >
> > >
> > > The box is stable, but performs terribly. Surely, I'm doing something
> > > wrong, but I would like some tips and tricks to speed things up.
> > >
> > >
> > >
> > >
> > > Here's my setup:
> > > CPU: Intel(R) Xeon(R) CPU E5-2620 v2 @ 2.10GHz (HyperThreading is
> > > enabled)
> > > RAM: 64GB
> > > NIC: 2x igb in lagg0 (loadbalancing)
> > Oops, I didn't see this before my last post. igb had problems with
> > the 64K TSO issue and I'd try to get rid of lagg as well.
> > (You might be much better off just using a single net interface
> >  without lagg.)
> >
> > Again, good luck with it, rick
> >
> > > Disks:
> > >
> > > export1     1.81T   914G   942G    49%  1.00x  ONLINE  -
> > >   mirror     928G   457G   471G         -
> > >     da0         -      -      -         -
> > >     da1         -      -      -         -
> > >   mirror     928G   457G   471G         -
> > >     da2         -      -      -         -
> > >     da3         -      -      -         -
> > >   mirror    9.94G   173M  9.77G         -
> > >     da4p1       -      -      -         -
> > >     da5p1       -      -      -         -
> > > cache           -      -      -      -      -      -
> > >   da4p2      223G   223G     8M         -
> > >   da5p2      223G   223G     8M         -
> > >
> > >
> > > da0-3 are 1TB WDs
> > > da4-5 are 240GB Samsung SSD 840s
> > >
> > >
> > > Here's (related) info from rc.conf.
> > >
> > > nfs_server_enable="YES"
> > > nfs_server_flags="-u -t -n 128"
> > > rpcbind_enable="YES"
> > > mountd_enable="YES"
> > > rpc_lockd_enable="YES"
> > > rpc_statd_enable="YES"
> > >
> > >
> > >
> > >
> > > I have compression enabled on all the ZFS-filesystems, and
> > > jumboframes are enabled on the nics.
> > >
> > >
> > > As soon as one of the (Linux) clients start to do some IO, NFS
> > > responsetimes go up bigtime (yesterday up to 13 seconds), while the
> > > hardware is pretty much idle, I must be doing something very wrong.
> > > I'm mostly a Linux-guy, so any hit with a FreeBSD cluebat is
> > > appreciated.
> > >
> > >
> > > Regards,
> > >
> > >
> > > --
> > > Kerio Operator in de Cloud? https://www.kerioindecloud.nl/
> > > Mark Schouten  | Tuxis Internet Engineering
> > > KvK: 61527076 | http://www.tuxis.nl/
> > > T: 0318 200208 | info at tuxis.nl
> >
> 
> 
> 
> A terrible write performance may come from Linux mount options in fstab :
> 
> 
> There is sync / async for the NFS connection :
> 
> I was set sync : A file requiring 30 ( thirty ) SECONDs to write become 30
> MINUTEs to write .
> After working to solve this problems by trying diffrerent parameter setting
> over many days , there only remained sync / async .
> When I selected async , writes turned to NORMAL .
> 
> I do not know effect of compression ( my opinion is that it will not be
> "terrible" ) , but my suggestion is to check as a possible trouble point :
> 
> sync / async in Linux computer ( in fstab or mount statement )

This sounds like you might want a ZIL added to your server...  Make
sure you use an SSD..  Since you're using an NFS server, it cannot
reply success to an operation till it is committed to stable storage
which it sounds like is slow...  With a ZIL on an SSD, it can more
quickly commit the NFS request to stable storage and return success
much quicker...

-- 
  John-Mark Gurney				Voice: +1 415 225 5579

     "All that I will do, has been done, All that I have, has not."