[Linux-cluster] Cluster with shared storage on low budget

[Gordan Bobic]

Nikola Savic wrote:
> Digimer wrote:
>> Once, and *only* if the fence was successful, the cluster will reform.
>> Once the cluster configuration is in place, recovery of the file system
>> can begin (ie: the journal can be replayed). Finally, normal operation
>> can continue, albeit with one less node. This is also where the resource
>> manager (rgmanager or pacemaker) start shuffling around any resources
>> that were lost when the node went down.
>>   
> 
>   From guide you sent me, I understood that fencing to work well servers
> should have IPMI available on motherboards.
> 
>   My client is going to purchase servers at Hetzner from their EQ-Line.
> I asked their support if IPMI is available. Since my other client
> already has server with 'em, I tried to install ipmi related packages
> (like you specified in guide). IPMI service doesn't start, so I assume
> it's not available or not turned on in BIOS.

That doesn't mean much. The IPMI service isn't what you use for fencing 
in this context. It's for diagnostics (e.g. advanced sensor readings, 
fan speeds, temperatures, voltages, etc.). Think of it as lm_sensors on 
steroids. For fencing you need to connect to the machine externally over 
the network via IPMI, and this will run at firmware level (i.e. you need 
to be able to power the machine on and off without an OS running).

>   How would cluster work if no IPMI or similar technology is available
> for fencing? In case one of nodes dies and no fencing is available,
> cluster will hang until administrator does manual fancing?

Yes, that's about the size of it.

There are add-in cards you can use to add fencing functionality even if 
you don't have this built into the server, e.g. Raritan eRIC G4 and 
similar. I wrote a fencing agent for those, you should be able to find 
it in the redhat bugzilla. They can be found for about £175 or so. That 
may or may not compare favourably to what you can get with the servers 
from the vendor.

Alternatively, you can use network controllable power bars for fencing, 
they may work out cheaper (you need one eRIC card per server, and 
assuming your servers have dual PSUs, you'd only need two power bars).

Something else just occurs to me - you mentioned MySQL. You do realize 
that the performance of it will be attrocious on a shared cluster file 
system (ANY shared cluster file system), right? Unless you only intend 
to run mysqld on a single node at a time (in which case there's no point 
in putting it on a cluster file system).

Gordan