Each CPS server is analogous to a single coordinator disk. A single CPS server will not provide the correct number of coordination points (minimum of 3) needed in a fencing environment. That is why you would need 3 CPS servers in a pure CPS environment that did not have any disk-based coordination points. Part of the value proposition is that the CPS servers could be located anywhere and that the physical location was not important as long as there was network connectivity. So in a stretch cluster that goes between sites A and B, I could locate one CPS server at each site, and for the tie-breaker I would locate the 3rd CPS server at C, provided that all 3 sites were connected on the network.
The Coordination Point Server will support either a pure CPS environment where you have a CPS server for each of your coordination points and there would be no SCSI disks used as coordination points, or you can have a mixed environment with both SCSI disks and CPS servers working together.
In either case you will want an odd number of total coordination points (disk-based or server-based). So you could have 3 CPS servers (or more). You could have 3 coordinator disks. You could have 2 coordinator disks and 1 CPS server. You could have 1 coordinator disk and 2 CPS servers.
The odd number is used to insure that in a partitioned cluster, one of the partitions will be able to win more coordination points than the other partition. When a cluster experiences a network partition, typically 2 sub-clusters remain, so we would need at least 2 coordination points, plus one more to insure that one of the sub-clusters can win the race for the keys.
