Gluster users - Nov 2016 - Improving IOPS

Il 06/11/2016 03:37, David Gossage ha scritto:
> The only thing you gain with raidz1 I think is maybe more usable 
> space. Performance in general will not be as good, and whether the 
> vdev is mirrored or z1 neither can survive 2 drives failing.  In most 
> cases the z10 will rebuild faster with less impact during rebuild. If 
> you are already using gluster 3 node replicate as VM practices suggest 
> then you are already pretty well protected if you lose the wrong 2 
> drives as well.
Ok, i'll try again. I'm *not* talking about a single RAIDZ1 for the 
whole server.

Let's assume a 12 disks server. 4TB each. Raw space = 4TB*12 = 48TB

You can do one of the following:
1) *a single RAIDZ10*, using all disks, made up by 6 RAIDZ1 mirrors. 
usable space=4TB*6 = 24TB
2) *6 RAIDZ1 mirrors*. usable space=4TB*6 = 24TB

You'll get the same usable space for both solution.

Now you have gluster, so you have at least 2 more servers in
"identical"
configuration.

With solution 1, you can loose only 1 disk for each pair. If you loose 2 
disks from the same pair, you loose the whole RAIDZ10 and you have
to heal 24TB from the network.

With solution 2, you can loose the same number of disks, but if you 
loose 1 mirror at once, you only have to heal that mirror from the 
network, only 4TB.

* IOPS should be the same, as Gluster will 'aggragate' each pair in a 
single volume, like a RAID10 does, but you get much more speed during an 
healing.
* Resilvering time is the same, as ZFS has to resilver only the failed 
disk with both solutions.

What i'm saying is to skip the "RAID0" part and use gluster as 
aggragator. Is much more secure and faster to recover in case of 
multiple failures.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://www.gluster.org/pipermail/gluster-users/attachments/20161106/06246486/attachment.html>

On Sun, Nov 6, 2016 at 3:24 AM, Gandalf Corvotempesta <
gandalf.corvotempesta at gmail.com> wrote:
> Il 06/11/2016 03:37, David Gossage ha scritto:
>
> The only thing you gain with raidz1 I think is maybe more usable space.
> Performance in general will not be as good, and whether the vdev is
> mirrored or z1 neither can survive 2 drives failing.  In most cases the z10
> will rebuild faster with less impact during rebuild. If you are already
> using gluster 3 node replicate as VM practices suggest then you are already
> pretty well protected if you lose the wrong 2 drives as well.
>
>
> Ok, i'll try again. I'm *not* talking about a single RAIDZ1 for the
whole
> server.
>
> Let's assume a 12 disks server. 4TB each. Raw space = 4TB*12 = 48TB
>
> You can do one of the following:
> 1) *a single RAIDZ10*, using all disks, made up by 6 RAIDZ1 mirrors.
> usable space=4TB*6 = 24TB
> 2) *6 RAIDZ1 mirrors*. usable space=4TB*6 = 24TB
>
I see maybe you don't really means raidz1 here.   Raidz1 usually refers to
"raid5" type vdevs with at least 3 disks otherwise why pay a penalty
for
tracking parity when you can have a mirrored pair.  So in your case you are
changing it from one zpool like was laid out to multiple zpools with each
one being 1 mirrored vdev pair of disks?

tank1
   mirror
    pair-a
    pair-a

tank2
   mirror
    pair-b
    pair-b

etc.....

as opposed to

tank1
  mirror
     pair-a
     pair-a
   mirror
     pair-b
     pair-b
>
> You'll get the same usable space for both solution.
>
> Now you have gluster, so you have at least 2 more servers in
"identical"
> configuration.
>
> With solution 1, you can loose only 1 disk for each pair. If you loose 2
> disks from the same pair, you loose the whole RAIDZ10 and you have
> to heal 24TB from the network.
>
> With solution 2, you can loose the same number of disks, but if you loose
> 1 mirror at once, you only have to heal that mirror from the network, only
> 4TB.
>
> * IOPS should be the same, as Gluster will 'aggragate' each pair in
a
> single volume, like a RAID10 does, but you get much more speed during an
> healing.
> * Resilvering time is the same, as ZFS has to resilver only the failed
> disk with both solutions.
>
> What i'm saying is to skip the "RAID0" part and use gluster
as aggragator.
> Is much more secure and faster to recover in case of multiple failures.
>
So moving from a replicated to a distributed-replicated model?  or a
striped-distributed-replicate?  what is the command or layout you would use
to get to the model you are wanting to use?
>
> _______________________________________________
> Gluster-users mailing list
> Gluster-users at gluster.org
> http://www.gluster.org/mailman/listinfo/gluster-users
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://www.gluster.org/pipermail/gluster-users/attachments/20161106/c466f407/attachment.html>