When planning a new virtualization environment, consolidation numbers are always flying around, and specifically the number of VMs you can run on a host.
According to This Doc you can have a max of 320 VMs per host, but keep in mind the number’s different for HA clusters. I was also pleased to find out the numbers were slightly changed for Update 1.
- 320 VMs per host
- 160 VMs per host in an HA cluster up to 8 hosts (100 if you’re on 4.0, non-Update 1)
- 40 VMs per host in an HA cluster of over 8 hosts
- Limit of 32 hosts per HA cluster
I’m ignoring non-HA environments for obvious reasons. So look at it this way:
- 8 hosts in an HA cluster can support up to 1280 VMs; 1 host reserved for failover would yield 1120 (12.5% overhead)
- 32 hosts in an HA cluster can support up to 1280 VMs; 1 host reserved for failover would yield 1240 (3.125% overhead) – Personally, I’d reserve 4 nodes in such a large cluster, which is 12.5% overhead and still yields the same 1120 VMs as the 8-node cluster
Doing that math clearly shows that anything bigger than an 8-node cluster is wasteful. “But I have 10 hosts I want to use, now what?” Build two 5-node clusters, one 10-node cluster supports up to 400 VMs, but two 5-node clusters with 1 reserved host each (leaving 8 usable nodes), will support up to 1280 VMs.
I originally wanted to build a 32-node cluster, but have since changed my mind and am going to build four 8-node clusters. Staying with the 12.5% overhead, that’s 4,480 VMs across those 32-nodes, where I’d be limited to a maximum of 1280 in a single 32-node cluster.
Granted, most hosts probably couldn’t (or shouldn’t) handle that many VMs, but with the future Dell hardware I’ve seen (which I can’t describe due to NDA), RAM won’t be your limitation. Also, from what they’ve shown me, adding additional DDR3 sticks doesn’t slow you down like previous Nehalem, Westmere appears to stay at 1066 (although I may be wrong here, since I haven’t seen all configurations available yet). I’ll just say that the 1TB RAM max in vSphere will be obtainable fairly easily. One whole terabyte in a single 4U server? That’s just crazy :D