HAST considarations

[George Kontostanos]

On Mon, Feb 6, 2012 at 1:51 PM, Daniel Kalchev <daniel at digsys.bg> wrote:
>
>
> On 06.02.12 13:43, Damien Fleuriot wrote:
>>
>> This issue is due to a bug in OpenBSD 3.8's implementation of CARP.
>>
>> It triggers if you have net.inet.carp.preempt=1 on the node.
>>
>> If the sysctl is set, the interface assumes MASTERship immediately upon
>> being brought up, then yields in the presence of a better master.
>
>
> I know about this patch, but on my systems
>
> net.inet.carp.preempt=0
>
> I was running 8-stable, now 9-stable on these servers and observe the same
> behavior.
>
> George Kontostanos:
>
> My setup has 1G interfaces for the CARP/Internet and 10G interfaces for the
> backend/HAST. I am doing hast over the 10G interfaces. For a system with
> part of 10k SAS drives, and ZFS mirror (each element of the mirror is an
> HAST provider), running bonnie++ I see about 100MB/sec flow to the secondary
> HAST and that about saturates the disks as well (50-60MB/sec.. should have
> been better)
> I had earlier experiment with 4 drives in each system and that replicated at
> up to 230 MB/sec.
>
> Daniel

I wonder how that numbers would look on striped raidz1 with 10 disks
with separate mirrored cache devices.

Another question that comes to my mind right now is if the failover
scripts should bring up/down the required services for data sharing.
That is given the fact that the pool is not mounted until their role
is primary, samba for example would complain about it.

Also, how about backup strategies? My feeling is that for true HA the
only option would be to have a network backup solution that would take
over from the currently active node. Of course this requires the
failover scripts to start /stop this service as well.

Regards
-- 
George Kontostanos
Aicom telecoms ltd
http://www.aisecure.net