Thu Aug 20, 2009 3:29 pm
I considered trying to do this but I could think of a lot of reasons why this would be a bad idea. So I'm waiting for Starwind HA instead.
If you are trying to achieve seamless failover, so that if windows/starwind fail for any reason, none of your iSCSI connections fail and are routed to another starwind box in the cluster, I don't think Windows failover clustering is cut out for this and you are better waiting for the HA features which the Starwind dev team are working on right now.
Last time I checked the Starwind license on this point (a long time ago!), you would need one for each node.
Why it's a bad idea...
Cost. Windows only supports iSCSI (with persistent reservation), Fiber Channel, and SAS (2 node only) as cluster storage. This will probably mean you have to invest in a costly hardware array and be tied to the array vendor for disks (see how much Dell want for a 5400rpm 1TB sata disk for an MD3000!).
Performance I assume you wouldn't want to use Starwind as an iSCSI front end to another iSCSI SAN. Almost every hardware iSCSI box (with controller redundancy etc) only has 1 to 4 gigabit Ethernet connections, so if you want 10GbE to your initiators, you will need a 8gbit/sec FC or 12gbit/sec SAS array to not be a bottleneck (in the real world it depends on how fast your drives are).
Reliability Microsoft failover clustering is primarily designed for application failover. The working nodes in a cluster notice that another node has failed, and restart that app on another node. There is no realtime sync of application state unless the application is specifically built for it or that state is stored on the shared storage. E.g. if a a hyper-v node fails, all the VMs on that node crash. The cluster "reboots" them on other nodes.It's only with planned failovers that you can move a VM from one node to another without shutting it down. I've never seen them claim that they can do seamless failover for block access (e.g. iSCSI) with their own iSCSI server, which is clusterable. I think Starwind would face similar issues. If Microsoft failover clustering was suitable for this, then I guess Starwind would be using it to build their HA features, rather than rolling their own solution.
Another option: NEC (i think OEMed from someone else) used to do / may still do redundant servers. This is like RAID-1 and redundant PSUs taken to the extreme - hardware mirrors CPUs and RAM in realtime. The OS has no idea. They were expensive! But did give you almost total hardware redundancy (a software crash would still take down your system). Vmware are supposed to be working on a software version of this.
