VM Density & Host Sizing

What it works out

Consolidation is arithmetic, and the interesting part is which constraint binds first. A database estate reaches the memory ceiling with cores to spare; a compute estate fills its cores with memory half empty. The tool names which, because that is what tells you what to change.

• Total vCPU, memory and storage across every configuration
• Physical cores after the overcommit ratio, memory at 1:1
• Hosts required by CPU and by memory separately, and which binds
• Per-host shape, spread evenly rather than filled to the brim
• Failover capacity for N+1 or N+2, and what share of the cluster it costs

Why memory has no overcommit setting

CPU overcommit degrades under load. Memory overcommit takes an environment down: ballooning and swapping fail non-linearly and the damage lands on whichever machine is busiest. Memory is sized 1:1 here and the option to do otherwise is not offered.

What high availability actually requires

A spare host is not high availability. It needs storage every host can reach, a migration network fast enough that evacuation stays practical, heartbeat traffic on its own path, and no single switch or power feed in the way. All four are returned when HA is selected.

Consolidation is a CPU decision and a memory rule

Virtual machines share physical cores well. A vCPU is a scheduling entity, and most workloads spend most of their time idle, so two or three vCPU per physical core is routine for application estates and comfortable for VDI. Databases and latency-sensitive systems are the exception, and 1:1 is the honest setting for them.

Memory does not work that way, and the difference is not a matter of degree. A vCPU waiting for a core is slow; a virtual machine that cannot get the memory it was promised is a machine the hypervisor starts taking pages from, and the failure mode is an outage rather than a slowdown. This calculator has no memory overcommit setting, and that is deliberate.

• 1:1 — no consolidation. Databases, real-time systems, anything with a latency SLA
• 1:2 — the common setting for mixed application estates
• 1:3 — VDI and bursty workloads, where most desktops are idle most of the time

Two ceilings, and only one of them binds

A host has a maximum core count and a maximum memory capacity, and an environment runs into one of them before the other. A database estate reaches the memory ceiling with cores to spare — those hosts are bought for RAM. A compute estate does the opposite and fills its cores while memory sits half empty.

Knowing which one binds tells you what to do about it. Memory-bound means a larger memory ceiling per host cuts the host count directly. CPU-bound means either bigger processors or, if the workloads genuinely tolerate it, a higher overcommit ratio. Sizing that reports only a host count hides the lever.

High availability is failover capacity, not a spare box

N+1 does not mean buying one more server and leaving it switched off. It means the cluster is sized so that when any one host fails, the survivors have the capacity to take everything it was running — the load spreads across fewer machines, and those machines must have been specified for that moment rather than for a normal Tuesday.

The useful consequence is that HA gets proportionally cheaper as the cluster grows. Protecting a three-host cluster costs a third of it. Protecting a twenty-host cluster costs five percent, for the same protection. A small cluster that needs HA is often better served by making the hosts smaller and buying more of them.

The requirement people discover last

High availability needs storage every host can reach. A virtual machine cannot restart on a surviving host if its disk only exists on the one that failed — so HA implies a SAN, an iSCSI or NFS array, or a hyper-converged layer spread across the hosts. That decision shapes the network, the budget and the failure behaviour of the whole cluster, and it is routinely left until after the servers are chosen.

The network matters as much. Live migration moves a host's worth of memory across the wire, which on a shared gigabit link takes long enough that people stop evacuating hosts before maintenance — quietly removing the thing HA was bought for. Heartbeat traffic needs its own path, because a burst of backup traffic that delays a heartbeat looks exactly like a dead host, and the cluster will fence a machine that was working perfectly.