8:34 p.m.
On Dec 24, 2019 02:18, Jayme <jaymef(a)gmail.com> wrote:
If you can afford it I would definitely do raid. Being able to monitor and replace disks
at the raid level is much easier than brick. With raid I’d do a gluster arbiter setup so
your aren’t losing too much space.
Keep an eye on libgfapi. It’s not default setting due to a few bugs but I’ve been testing
it in my ssd hci cluster and have been seeing up to 5x io improvement. Also jumbo frames
on those 10Gbps switches.
Someone else will probably chime in re the other questions. I believe the GUI can only
deploy a three sever cluster then you have to add the remaining hosts afterward.
On Mon, Dec 23, 2019 at 5:56 PM <adrianquintero(a)gmail.com> wrote:
>
> Hi,
> After playing a bit with oVirt and Gluster in our pre-production environment for the
last year, we have decided to move forward with a our production design using ovirt 4.3.7
+ Gluster in a hyperconverged setup.
>
> For this we are looking get answers to a few questions that will help out with our
design and eventually lead to our production deployment phase:
>
> Current HW specs (total servers = 18):
> 1.- Server type: DL380 GEN 9
> 2.- Memory: 256GB
> 3.-Disk QTY per hypervisor:
> - 2x600GB SAS (RAID 0) for the OS
I would pick Raid1 and split that between
OS and brick for engine 's gluster volume
> - 9x1.2TB SSD (RAID[0, 6..]..? ) for GLUSTER.
If you
had bigger NICs I would preffer RAID0. Have you considered JBOD approach (each disk is
presented as a standalone LUN from the raid array) ?
Raid 6 is waste for your SSDs (when you already have replica 3 volumes). Ask HPE to
provide SSDs from different manufacturing batches. I think that with SSD thee best
compromise will be RAID 5 with replica3 for faster reads.
As the SSDs will be overwritten with the same ammount of data, there could be a situation
where all 3 from a replica set will be in predictive failiure (which is not nice) -
that's why a raid 5 is a good option.
> 4.-Network:
> - Bond1: 2x10Gbps
> - Bond2: 2x10Gbps (for migration and gluster)
Do not use Active-Backup bond as you will loose bandwidth. LACP (with hashing on layer 3+
layer4) is a good option, but it depends if your machines will be in the same LAN segment
(communicating over layer 2 and no gateway inbetween).
> Our plan is to build 2x9 node clusters, however the following
questions come up:
>
> 1.-Should we build 2 separate environments each with its own engine? or should we do
1 engine that will manage both clusters?
If you decide to split - you have less fault tollerance - for example loosing 3 ot of 18
is better than loosing 3 out of 9 nodes :)
> 2.-What would be the best gluster volume layout for #1 above with
regards to RAID configuration:
> - JBOD or RAID6 or…?.
RAID6 wastes alot of space
> - what is the benefit or downside of using JBOD vs RAID 6 for
this particular scenario?
> 3.-Would you recommend Ansible-based deployment (if supported)? If yes where would I
find the documentation for it? or should we just deploy using the UI?.
There is a
topic in the mailing list about issues with ansible.It's up to you.
> - I have reviewed the following and in Chapter 5 it only mentions
Web UI
https://access.redhat.com/documentation/en-us/red_hat_hyperconverged_infr...
> - Also looked at
https://github.com/gluster/gluster-ansible/tree/master/playbooks/hc-ansib... but
could not get it to work properly.
>
> 4.-what is the recommended max server qty in a hyperconverged setup with gluster, 12,
15, 18...?
I don't remmember but I think it was 9 nodes for Hyperconverged setup
in RHV.
> Thanks,
>
> Adrian
Best Regards,
Strahil Nikolov
1795
days inactive
1795
days old
0 comments
1 participants
participants (1)
-
Strahil