On Mon, 2007-01-08 at 12:18 -0200, Marco Sinhoreli wrote:
> Hello everybody,
>
> We are thinking about using virtual machines for a scene of high
> availability where is contemplated the division of the VMs in two hosts.
> Considering that both hosts will have 6 VMs and in the host A will be
> active the VMs vm1, vm3 and vm5 , while in the host B will be active
> the VMs vm2, vm4 and vm6. For each VBD allocated in one LV (Logical
> volume of the LVM) we pretend to implement the following architecture.
>
> P = primary | S = Secondary
> ------------------------------------
> Host A
> ------------------------------------
> vm1 vm2 vm3 vm4 vm5 vm6
> | | | | | |
> drbd1 drbd2 drbd3 drbd4 drbd5 drbd6
> P S P S P S
> ------------------------------------
> | | | | | |
> | | NETWORK | |
> | | | | | |
> ------------------------------------
> S P S P S P
> drbd1 drbd2 drbd3 drbd4 drbd5 drbd6
> | | | | | |
> vm1 vm2 vm3 vm4 vm5 vm6
> ------------------------------------
> Host B
> ------------------------------------
>
> Consider that, where the block device is primary, the VM is active on
> the host, and where the device blocks is secondary, the VM is inactive
> on this host.
You may find this is easier utilizing network attached (centralized)
storage in lieu of block level mirroring, however your explanation is
somewhat confusing.
>From your block diagram, it looks as though you plan to live-migrate in
the event of failure or scheduled shutdown, using nbd to maintain the
file system for each of the 4 components making up host A or B, each
having separate file systems. Is this correct?
>
> The heartbeat will have one responsible script for the VMs allocation in
> case of hardware failure from one of the hosts for your even. In case of
> programmed stop, the host from the VMs will do the migration for it's
> even without the stop of the services that are been executed. Will occur
> a prolongated downtime and full lost of the VMs activities only in case
> of a abrupt stop of the servers hardware.
So at no point will two active VM's be accessing the same file system,
correct? This would be easier if the file systems were on a raid backed
NAS.
>
> We wish information from the list if this solution have been done by
> someone,in positive case, if there is some documentation that
> contemplate this implementation ,and in negative case, how to do an
> implementation of a HA integrated with Xen.
Your solution should work exactly as you described it, if I'm
understanding it correctly. My only recommendation is to ditch block
level mirroring and go with centralized storage.
You can, of course, use NBD to mirror the storage if a decent RAID
doesn't give you the desired comfort level. What you have , in essence
is four single pole double throw switches (an over simplified example),
ensuring there is no "short" (no two P/S domains access the same FS at
once) and you're fine :)
A little over paranoid considering I/O costs, but should work well :)
>
>
> Best regards
>
>
> Marco Sinhoreli
> Linux expecialist
> Samurai Projetos Especiais
> São Paulo - Brazil
>
>
Best,
--Tim
> _______________________________________________
> Xen-users mailing list
> Xen-users@xxxxxxxxxxxxxxxxxxx
> http://lists.xensource.com/xen-users
_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-users
|
Home