Windows Server 2016 shipped last week, and we spent the end of September in Atlanta at Ignite watching Microsoft launch it. We have been running the Technical Previews in the lab since TP4, so GA is less a revelation than a starting gun: the question our clients are asking is not "what is new" — the feature lists are everywhere — but "which Hyper-V 2016 features justify touching a stable virtualization estate, and when."
That question has a concrete backdrop for us. Over the past several years we have virtualized more than 400 servers onto Hyper-V for clients ranging from government agencies to healthcare groups, with full System Center management on top. Those clusters run Windows Server 2012 R2, they are boring in the best way, and every one of those clients now wants an opinion on 2016. This is the opinion, feature by feature, judged against real private cloud workloads rather than demo scenarios.
Short version up front: the licensing change is the most important slide from Ignite for most budgets, Storage Spaces Direct is the feature we are most excited about and least in a hurry to deploy, and two small features nobody keynotes — PowerShell Direct and rolling cluster upgrades — will improve our operational life more than anything else this year.
The per-core licensing change is the real headline
Windows Server 2016 moves from per-socket to per-core licensing: all physical cores must be licensed, minimum 8 cores per processor and 16 per server. For a two-socket host with 8-core processors, the cost works out equivalent to 2012 R2. For the two-socket, 18-core-per-processor hosts that our denser clusters run — the ones we specified precisely because per-socket licensing rewarded core density — the Datacenter edition math changes materially upward.
This lands directly on private cloud economics. The consolidation playbook of the last five years said: buy the fattest cores-per-socket you can, stack VMs deep, and licensing stays flat. That playbook is now partially repealed. For clients planning 2017 host refreshes, we are re-running the hardware sizing with licensing as a first-class input again — in some cases a four-socket-worth of moderate-core hosts beats two monster hosts once Datacenter core packs are priced in. If you take one action this quarter, inventory your physical core counts before your renewal, not after.
Storage Spaces Direct: the feature we want, deployed the year we'll trust it
Storage Spaces Direct (S2D) is the big architectural swing: local disks in each cluster node pooled into shared storage, no SAN, no shared JBOD, with the Software Storage Bus and RDMA networking doing the work a fibre channel fabric used to do. Combined with Storage Replica for cluster-to-cluster replication, Microsoft is describing a private cloud storage stack with no storage vendor in it.
We believe the architecture. We have run it in the lab on a four-node cluster since TP5, and the converged and hyper-converged modes both behave as advertised; ReFS with accelerated VHDX operations makes checkpoint and fixed-disk creation almost disconcertingly fast. Our hesitation is not about whether it works — it is about operational maturity around failure. A SAN vendor's array has fifteen years of scar tissue about what happens when a disk lies about a write. S2D at GA has strict hardware compatibility requirements (watch the HCL discipline on drives and NICs; RDMA misconfiguration is the new "it's always DNS"), and the field-diagnostics tooling is version one.
Our plan, and our advice: deploy S2D in 2017 for the workloads where its economics are overwhelming — VDI storage tiers, test/dev clusters, branch-scale deployments — and let the production tier-1 SQL clusters watch from their SAN for another cycle. The 35% infrastructure cost reductions we have delivered for clients came from being aggressive about consolidation and honest about risk sequencing. Same discipline here.
Shielded VMs: the government conversation just changed
Shielded VMs — generation 2 VMs with virtual TPM, BitLocker-encrypted disks, and a Host Guardian Service that attests host health before releasing keys — solve a problem most commercial clients have never articulated but every government security officer articulates constantly: the fabric administrator is in the trust boundary. Today, anyone who can administer the Hyper-V host can copy a VHDX and mount it elsewhere. With shielding, a stolen or copied VM disk is ciphertext, and a compromised or non-attested host cannot start the VM at all.
For our state government clients, this is the first Hyper-V feature in years that changes an accreditation conversation rather than a performance one. It is also operationally serious: HGS is a new tier-0 service (we will run it as a separate three-node cluster in its own forest, per guidance), attestation mode matters (TPM-based attestation needs TPM 2.0 host hardware; the Active Directory-based mode is easier and weaker), and your recovery runbooks must account for the fact that losing HGS means your shielded VMs do not start. We are building the reference design in the lab now, with a pilot targeted at a small enclave of high-sensitivity VMs, not the general estate.
Nested virtualization and PowerShell Direct: small features, outsized field value
Nested virtualization — Hyper-V inside a Hyper-V VM — reads like a lab toy, and mostly it is one. But labs are where migrations succeed or fail. Every engagement we run includes a rehearsal environment, and until now that meant physical hosts dedicated to rehearsing cluster operations. Now a full multi-node "cluster," with failover behavior close enough for runbook rehearsal, fits inside VMs on one physical box. Our delivery cost for rehearsal environments just dropped, which means our clients' price for rehearsal just dropped, which means fewer clients will skip it. That is a safety improvement disguised as a convenience.
PowerShell Direct is the other quiet winner: a PowerShell session into a VM over the VMBus, no network path, no WinRM configuration, just host-level access and guest credentials:
Invoke-Command -VMName APP-VM-014 -Credential $guestCred -ScriptBlock (
Get-Content .\remediate-vm-network.ps1 -Raw | Invoke-Expression
)
Every virtualization operator knows the failure mode this addresses: the VM whose network configuration you broke is the VM you can no longer reach to fix. On a 400-server estate, that scenario shows up a few times a year and previously ended with someone driving to use a console session. Now it is a one-liner from the host.
The operational features that pay rent immediately
Three more items that will not headline anyone's keynote but map straight onto pain we log tickets for today:
- Cluster OS Rolling Upgrade. You can add 2016 nodes to a 2012 R2 cluster and upgrade node-by-node with the cluster functional level raised only at the end — no new cluster, no big-bang migration, VMs live-migrating around the work. This is the adoption feature: it converts "migrate 400 VMs to new clusters" into "patch cycle with extra steps." Note the one-way door: once you run
Update-ClusterFunctionalLevel, there is no going back to mixed mode. - Production checkpoints. Checkpoints that use VSS in the guest instead of saved memory state, making them safe for domain controllers and SQL in ways the old ones never were. Our standing "no checkpoints on production VMs without an architect sign-off" policy gets its first rewrite since 2012.
- Hot add memory and NICs, and VM configuration versions. Runtime resize removes a whole category of change windows. And because 2016 does not force VM configuration upgrades, you can keep VMs at version 5.0 and mobile back to 2012 R2 hosts until you are certain — upgrade the version deliberately, per VM, with
Update-VMVersion, and treat it as the commitment it is.
We are also keeping a skeptical eye on Nano Server as a Hyper-V host option — the density and attack-surface story is attractive, the operational retooling it demands is significant, and we would rather let the Current Branch for Business cadence shake out before betting host infrastructure on it.
Our adoption sequence for existing 2012 R2 estates
For clients on our managed clusters, the sequence we are proposing is: new clusters and refresh hardware get 2016 from day one, effective now that GA media is out. Existing clusters take the rolling upgrade path starting Q1 2017, after System Center 2016 is in place beneath them — SCVMM 2012 R2 cannot manage 2016 hosts properly, so management upgrades first (that is next month's post). Shielded VM and S2D pilots run in parallel, scoped narrow. Nothing about a stable 2012 R2 cluster is broken, and urgency should come from features you will actually use, license timing, and hardware lifecycle — not from a launch keynote, however good the demos were.
If you're facing this
If you run a Hyper-V estate and are trying to decide when 2016 earns its way in — or whether S2D can really take your storage bill apart — we have done this evaluation against 400+ virtualized servers and are happy to compare notes. Get in touch; we will bring the lab results and the licensing spreadsheet.