[Gluster-users] Production cluster planning

[David Gossage]

On Wed, Oct 5, 2016 at 2:14 PM, Gandalf Corvotempesta <
gandalf.corvotempesta at gmail.com> wrote:

> 2016-10-05 20:50 GMT+02:00 David Gossage <dgossage at carouselchecks.com>:
> > The mirrored slog will be useful.  Depending on what you put on the pool
> > l2arc may not get used much.  I removed mine as it got such a low hit
> rate
> > serving VM's.
>
> I'll use shards. The most accessed shard isn't cached in L2ARC?
>

L2ARC filled up some, but the usage hit ratios never hit above 8% so I just
freed it up fro the system to use as it would.  Haven;t noticed any
difference since doing that.

>
> I'll also other pools for plain file hosting (websites and so one). In
> this case an L2ARC would be usefull. I think.
> Is possible to share the same L2ARC from multiple ZFS pools ?
>

L2ARC is per pool I believe.

>
> > I also re-did my bricks from raidz2 to using mirrored pairs. I felt that
> > with 3 server redundancy the performance benefits from a more raid10 type
> > layout would be more useful than raidz2 for the pool.  I could see where
> > some might want more security though in their data
>
> So, you removed the RAIDZ2 and created multiple mirrors (like RAID-1) ?
> I can do that, but is really a waste of space. In a replica 3, your
> are replicating 6 times.
> Or did you create a RAID-10 ? I really hate RAID-10. If you loose 2
> disks from the same mirror
> (and in a 12 disks server this could be frequent) you loose the whole
> pool and rebuilding time from
> the network could be.............. HUGE. Try to rebuild 24TB from
> network. In a perfect world, this require
> at least 24.000/1.24 = 19.200 seconds. more than 5 hours by using the
> whole 10GB network only for healing.
>
> I feel more confortable with 2xRAIDZ-2 than a RAID10
>

That's fine.  For my VM case I currently only use around 1TB per host so
worst case scenario I found was if one server is down I could rebuild brick
overnight.  For your case that's not as easy.  Also for me I use 1TB disks
in my mirrors so if one drive goes down it's not that long to re-sync it on
replace.  Reduces my chance to have 2 disks in same mirror die at same time
a little bit.

>
> By using SLOG, the RAIDZ2 write penalty should be removed as gluster
> always write to SLOG (SSD) and to the RAIDZ-2 only in background,
> right?
> Is possible to change the ZIL flush timeout from 5 seconds to
> something bigger as i'm using SSD with power-loss protection ?
>

I believe you can though I haven't played with that.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.gluster.org/pipermail/gluster-users/attachments/20161005/6f23f661/attachment.html>