| 1 | # Windows presets |
| 2 | |
| 3 | Prepared images live in the VM library's `Presets` directory. Each published |
| 4 | directory is immutable: existing guests retain QCOW2 backing-file references |
| 5 | to it. Publish a new ID with `supersedes` to change the library default while |
| 6 | preserving older bases. Host paths come from the configured VM library root. |
| 7 | |
| 8 | The provisioning contract is [windows-provision.ps1](../guest/windows-provision.ps1). |
| 9 | Prepare an isolated template with its account named `sandbox`, then run that |
| 10 | script with `-Prepare` from elevated 64-bit Windows PowerShell. Preparation |
| 11 | randomizes the template password, removes clone identity and completion state, |
| 12 | and installs the startup task. Shut down normally before capturing through |
| 13 | `vms.save_preset` with `setup=windows-v4`. The dashboard deliberately rejects |
| 14 | ordinary Windows preset capture because it cannot establish preparation. |
| 15 | |
| 16 | Validate fresh clones with different account names before publication. Check |
| 17 | the personalized account, unchanged computer name, one boot, the `ready` marker, Explorer, |
| 18 | the guest capture process, and the actual dashboard screen and input. Compare |
| 19 | the published disk and NVRAM bytes with the tested candidate, and check that |
| 20 | the disk is standalone with `qemu-img info` and `qemu-img check`. Never replace |
| 21 | an existing base in place. |
| 22 | |
| 23 | New clones retain the template computer name. Account personalization uses one |
| 24 | sign-out/sign-in so the desktop inherits the selected username, without a second |
| 25 | OS boot. Each clone still gets a separate password, MAC address, and DHCP lease. |
| 26 | Published `windows-v3` bases keep their original rename-and-reboot seed contract. |
| 27 | |
| 28 | ## 2026-10-20 rollout |
| 29 | |
| 30 | The `windows-10-pro-v10` candidate was prepared and tested on HDD storage with |
| 31 | four vCPUs, 4 GiB RAM, and a 64 GiB disk. Both its base and each clone's writable |
| 32 | overlay stayed on HDD. The host's 150 GiB SSD read cache remained enabled. |
| 33 | |
| 34 | Preparation refreshed the 64-bit .NET Framework native-image cache with |
| 35 | `ngen update /force` and installed native images for the assemblies loaded by |
| 36 | Windows PowerShell, then reran the normal `-Prepare` flow. All assembly-install |
| 37 | commands and preparation completed successfully. See Microsoft's |
| 38 | [Ngen documentation](https://learn.microsoft.com/en-us/dotnet/framework/tools/ngen-exe-native-image-generator) |
| 39 | for regeneration after framework servicing. No global CLR switches were added. |
| 40 | |
| 41 | | Candidate | Fresh clone to ready | Account | |
| 42 | | --- | ---: | --- | |
| 43 | | v9 repeat | 455 s | fixture | |
| 44 | | v10 first | 206 s | fixture | |
| 45 | | v10 repeat | 161 s | fixture2 | |
| 46 | |
| 47 | Both v10 clones completed personalization and reboot, started Explorer and the |
| 48 | guest capture process, and received distinct computer names. These are rollout |
| 49 | observations with differing cache warmth, not a controlled measurement of |
| 50 | L2ARC's contribution. Synchronous writes still reach HDDs. |
| 51 | |
| 52 | The published v10 preset supersedes v4. v4 and earlier referenced bases remain |
| 53 | available to their existing guests. Reverting the default requires withdrawing |
| 54 | v10 from discovery while preserving its bytes wherever clones still reference |
| 55 | them; it does not require modifying existing guest disks. |
| 56 | |
| 57 | ## First-run timing after fleet retirement |
| 58 | |
| 59 | After [retiring the unused guests](vm-retirement.md), another disposable v10 |
| 60 | clone reached `ready` in 95.75 seconds with the same 4-vCPU/4-GiB configuration. |
| 61 | The host observed these phases by polling the provisioning markers: |
| 62 | |
| 63 | | Phase | Duration | |
| 64 | | --- | ---: | |
| 65 | | Create disk clone and start | 1.12 s | |
| 66 | | First boot through provisioning startup | 48.63 s | |
| 67 | | Apply account and computer identity | 2.08 s | |
| 68 | | Restart, log in, and start desktop capture | 43.92 s | |
| 69 | |
| 70 | The guest agent first answered at 33.10 seconds. Guest marker timestamps put |
| 71 | identity changes at 1.83 seconds; the host's polling rounds this to 2.08 seconds. |
| 72 | Two Kernel-General boot events confirmed the reboot. Explorer and guest capture |
| 73 | were running, and the production browser displayed the personalized desktop. |
| 74 | |
| 75 | This is a disk template, not a saved-memory resume. `Rename-Computer` in the |
| 76 | provisioning script is followed by a forced restart; the second boot enables |
| 77 | capture and waits for the desktop. Avoiding that second boot is the next |
| 78 | architectural opportunity. Cloning the disk itself is already inexpensive. |
| 79 | |
| 80 | By `ready`, QEMU recorded 474 flushes with 46.45 seconds of aggregate duration |
| 81 | (98 ms per flush), 2.24 GB read, and 105 MB written. These operation durations |
| 82 | can overlap and are not an additive wall-clock breakdown. The read cache does |
| 83 | not accelerate synchronous HDD writes. This quieter-host result also has |
| 84 | different cache warmth, so it does not isolate the benefit of fleet retirement. |
| 85 | |
| 86 | The restricted raw trace is |
| 87 | `/var/lib/snowglobe/retirements/vms-20261020/windows-boot-trace.json`. |
| 88 | The trace guest was shut down and removed afterward; the published preset and |
| 89 | personal Mac remained intact. |
| 90 | |
| 91 | ## One-boot rollout |
| 92 | |
| 93 | `windows-10-pro-v11` supersedes v10 and uses `windows-v4`: the computer name stays |
| 94 | unchanged, while the selected username and a fresh password are applied to the |
| 95 | existing account SID/profile. A targeted sign-out/sign-in refreshes the desktop |
| 96 | environment. The SID-bound capture task starts in the new interactive session. |
| 97 | Temporary `ForceAutoLogon` is removed on both success and failure. |
| 98 | |
| 99 | With four vCPUs, 4 GiB RAM and a 64 GiB HDD overlay, the candidate reached ready |
| 100 | in 60.10 seconds; a fresh clone of the published v11 reached ready in 59.88 |
| 101 | seconds. That is about 36 seconds faster than the quiet-host v10 trace above. |
| 102 | Both machines kept the template computer name and had one Kernel-General boot |
| 103 | event since creation. Interactive username, SID/profile, password authentication, |
| 104 | Explorer and capture session ownership, and removal of `ForceAutoLogon` were |
| 105 | checked. These are individual runs with differing cache warmth. |