ZFS performance help sought

[Mason Loring Bliss]

On Thu, Jan 21, 2016 at 11:17:05PM -0700, Warren Block wrote:

> I know there was some sort of ZFS resource exhaustion issue worked on
> recently, but I don't recally exactly when or what the symptoms were. It
> might only be in 10-stable so far.  Best to ask on the freebsd-fs mailing
> list.

Thank you. I'm subscribing to that now.


> It's conceivable that you are seeing the result of lock contention that
> is slowing down arc_get_data_buf(), for details see:
> https://www.fabiankeil.de/gehacktes/electrobsd/zfs-arc-tuning/

I will look at that and try the dtrace script soon. Thank you.


> Even if it's a different issue, you may be able to work around it by
> throttling the send/receive throughput with mbuffer or a similar tool.

Most of the time I schedule the bulk moves for hours where I wouldn't notice
the system being bogged down, so I'm more interested in identifying it than
working around it, but that's worth noting, especially given that there is a
very easy chokepoint to be had in the pipeline.


On Fri, Jan 22, 2016 at 08:54:52AM -0500, kpneal at pobox.com wrote:

> FWIW, I have a machine with a three way mirror, SAS 3Gb, 8GB RAM, and when
> mutt saves a roughly 3GB mailbox it pushes the machine into swap.

I'll move further correspondence to freebsd-fs once I've confirmed my
subscription there, but I'd be curious to know if you can suggest a simple
test case there I can reproduce. Is it one big mbox file, as opposed to
maildir? Are you deleting a mail from the middle, or doing some other
operation? On my end, I take a dataset with maybe 130G in it and I remove it
from back-ups, and then re-ship it to tickle the issue.


On Fri, Jan 22, 2016 at 08:39:17AM -0700, markham breitbach wrote:

> There were some ZFS threads on freebsd-performance a while back too.  It
> might be worth throwing the question out that way too.

Thank you. I will subscribe and look at archives there too.

-- 
Mason Loring Bliss          mason at blisses.org          Ewige Blumenkraft!
awake ? sleep : random() & 2 ? dream : sleep; -- Hamlet, Act III, Scene I