Servers for virtualization

Memory first, cores second

Guests consume memory whether they compute or not, so host sizing starts with the memory target and works backwards to the chassis that holds it. The deepest memory configuration in our rack range is the HPE ProLiant DL380 Gen11 listing — up to 8 TB of DDR5 across 32 slots — with the 2U class generally carrying the memory depth virtualisation wants. Core count then sets how far that memory can be oversubscribed.

Platform shapes the same decision twice: high-core AMD EPYC models (the R7615's EPYC 9654 configuration, the SR655) maximise threads per host, while Xeon models suit estates standardised on Intel tooling. Per-core licensing can flip the arithmetic — price both effects at the quote stage.

The listed virtualisation candidates

The Lenovo ThinkSystem SR650 V4 is positioned in our shop range for virtualisation and AI workloads, and the 2U mainstream — PowerEdge R7xx, ProLiant DL380, ThinkSystem SR650 — is where most host orders land. Storage for the guests usually leaves the host: the array pages cover the shared tier, keeping hosts stateless and replaceable.

What to send

• Guest count and average guest size, or the estate being consolidated.
• The memory target per host and the oversubscription you accept.
• Whether guest storage is local or on a shared array.

Licensing belongs in the hardware conversation

Per-core licensing is the virtualisation topic that most changes hardware choices and is most often raised last. The same guest estate can carry very different software costs on a 32-core host and a 96-core host, and the difference can exceed the price gap between the machines. None of this is a reason to avoid dense hosts — consolidation saves real money in power, space and administration — but the arithmetic must be done whole: hardware, licences and operations in one column.

Bring the software position into the inquiry. The desk cannot price your licences, but it can configure against them — socket counts, core counts and host counts chosen with the licensing model in view — once it knows the model exists. A virtualisation quote that never asked is a quote that guessed.

Failure domains and the maintenance window

Host count is also an availability decision. Two hosts at sixty percent load can absorb each other's guests for a maintenance window; four at ninety percent cannot absorb anything. The N+1 pattern — the estate sized so one host can be empty without pain — is the difference between patching as a calendar event and patching as an incident. It costs one host. Estates that skip it pay the same money later, in overtime and risk, without the receipt.

One last sizing habit

Revisit the estate's real utilisation quarterly and let the figures, not the folklore, size the next hosts. Virtualisation estates drift toward the shapes their first sizing assumed; measurement is how they escape their own history.