I85PREP
⌘ K
DMCP study materials
DMCP study resources

DMCP: Servers and Compute Infrastructure

DMCP · Work through your exam, one topic at a time.

All guide sections

Servers and Compute Infrastructure

The Servers and Compute Infrastructure domain checks whether you can turn a workload profile into a host configuration - cores, memory, expansion, power, and firmware - and defend the headroom you left. Expect sizing scenarios where one resource, almost always memory, is the binding constraint while CPU looks generously provisioned.

Sizing rules

  • Count physical cores, not threads, for licensing and for most consolidation ratios. SMT adds throughput but not cores.
  • CPU overcommit of 4:1 is common for general virtual workloads; memory overcommit is where performance dies, because ballooning and swapping appear under pressure. Tier-1 production designs typically run 1:1 memory or better.
  • NUMA: keep a VM's vCPU count within a single socket where possible, and size per-socket memory so the hypervisor does not split wide VMs across nodes.
  • Memory population follows the vendor rule - usually balanced across channels with identical DIMMs per channel. Unbalanced population silently costs bandwidth.

Platform and lifecycle

  • Set a firmware and BIOS baseline per model, record it, and update using the vendor's interdependency matrix; firmware and driver versions are coupled.
  • UEFI boot with Secure Boot and a TPM is the default posture for new builds; legacy BIOS blocks measured boot and modern attestation.
  • Redundant PSUs on separate power feeds, with the rack power and thermal budget checked before the bill of materials is signed.
  • N+1 at the chassis level means the remaining hosts carry the load at target performance, not merely stay powered on.

Worked example

60 production VMs averaging 4 vCPU and 24 GB RAM. Candidate host: 2 x 32-core CPUs, 1 TB RAM. CPU demand = 240 vCPU; at 4:1 that is 60 physical cores, so one host suffices on CPU. Memory demand = 60 x 24 = 1,440 GB, so two hosts are needed on memory, and N+1 requires a third. Answer: three hosts, constrained by memory. The reasoning that matters is testing each resource independently and letting the largest result win.

Traps

  • Sizing on threads and then applying a core-based licence.
  • Assuming memory overcommit is safe because CPU overcommit is.
  • Forgetting hypervisor overhead (roughly 2-8 GB per host) and the N+1 host.
  • Mixing DIMM sizes across a channel to reach a capacity target.