Publishing a RemoteApp on a Windows VPS looks like a five-minute job: install the RDS role, create a collection, publish a program, send users a .rdp file. The part that bites is what happens at user number twelve. The host that felt instant with three test users starts swapping, logons stretch to ninety seconds, and the licensing grace timer quietly starts counting down. This article is about the two numbers nobody plans for: how many users fit, and what the licensing actually requires before you scale past them.
RemoteApp is still a session host
Every RemoteApp user gets a real session on a real session host. They see one window instead of a desktop, but Explorer, the profile, the logon process, and the per-session memory overhead all still happen. That means RemoteApp density is a session-host capacity question, not a “lighter tier” question. If you have not sized a session host before, the per-session memory budgets in the Windows VPS sizing guides here are the right starting reference.
Licensing before scaling
Get this wrong and the host stops accepting connections 120 days after install, which is a miserable way to discover a licensing gap. Three facts matter:
- You need an RD Licensing server with RDS CALs installed, and each session host must point at it. The grace period is 120 days, then connections are refused outright.
- RemoteApp does not avoid CALs. Published apps consume the same per-user or per-device RDS CAL as a full desktop session. Anyone claiming otherwise is confusing RemoteApp with a web app behind IIS.
- Per-user vs per-device changes your bill and your audit. Per-user wins when people use multiple devices; per-device wins for shift workers on shared thin clients. Mixed-mode is not allowed on one host.
# Confirm the host knows where its licences live
Get-ItemProperty "HKLM:\SYSTEM\CurrentControlSet\Control\Terminal Server\RCM\Licensing Core" |
Select-Object LicensingMode, LicenseServers
# Check how many CALs are issued vs available on the licensing server
Get-WinEvent -LogName "Microsoft-Windows-TerminalServices-Licensing/Operational" -MaxEvents 50 |
Where-Object Id -in 4105,4106,4107 | Select-Object TimeCreated, Id, Message
Capacity per vCPU and per GB
Published apps are lighter per session than full desktops because Explorer and its shell extensions are mostly out of the picture. A single business application in its own window typically holds 250–600 MB, against 1–1.5 GB for a desktop session running Office and a browser. The table below is a planning starting point for one application published to all users.
| Host size | Realistic RemoteApp users | Bottleneck you will hit first |
|---|---|---|
| 2 vCPU / 8 GB | 8–12 | Memory, then CPU during simultaneous logon |
| 4 vCPU / 16 GB | 18–25 | Logon storms and profile load, not steady-state CPU |
| 8 vCPU / 32 GB | 35–50 | Disk IOPS on profile writes and app temp files |
| 16 vCPU / 64 GB | 70–90 | Single-host blast radius — split before you get here |
Two adjustments move these numbers materially. An application that renders or computes heavily cuts density by a third or more regardless of RAM. And the moment you publish a second app to the same users, they stop closing the first one — density drops toward desktop-session numbers. Plan for the app set, not the single app you tested with.
The logon storm is the real ceiling
Steady-state CPU on a well-sized RemoteApp host is usually unremarkable. The peak arrives between 08:45 and 09:15 when forty people log in within ten minutes. Each logon loads a profile, applies Group Policy, starts the published app, and touches disk hard. Watch processor queue length and disk queue length during that window, not at 11am.
Mitigations that work, in order of effort: pre-load Group Policy with Always wait for the network at computer startup disabled, disable unused startup items in the published app, move profiles off the session host, and stagger start times by team. Only then consider more vCPU.
Profiles: where RemoteApp density quietly dies
With three test users, local profiles are fine. With thirty, local profiles are the reason your disk is saturated at 9am and why nobody can log in after a host rebuild. A RemoteApp user still gets a full user profile; the only question is whether it lives on the host or on a share.
| Profile approach | Good for | Watch out for |
|---|---|---|
| Local profiles | Under 8 users, single host | Disk growth, no portability, slow first logon |
| User Profile Disks (UPD) | Single collection, modest user count | VHDX per user consumes storage even when idle; no multi-collection sharing |
| FSLogix containers | Multiple collections or hosts, Office in the app stack | Needs a reliable SMB share and its own sizing |
If you are publishing a single line-of-business app to a stable team on one host, UPD is the pragmatic choice and takes twenty minutes to enable. If you expect a second host within a year, start with FSLogix now — migrating profile technology later is far more disruptive than setting it up correctly the first time.
A sizing checklist you can actually run
- Publish the app to three test users and measure per-session working set with
Get-Processfiltered to their session IDs. - Multiply by the real user count, add 2 GB for the OS and 1 GB for RDS role services.
- Check the licensing mode and confirm CALs are installed before the grace timer matters.
- Set idle and disconnected session limits so abandoned sessions release memory.
- Move profiles off the host before you exceed about ten users.
- Test a simulated morning logon with a script that opens ten sessions at once, and watch disk queue length.
# Per-session memory and process count, usernames resolved
Get-Process | Where-Object SessionId -gt 0 |
Group-Object SessionId |
Select-Object Name, Count, @{n='MB';e={[math]::Round((($_.Group | Measure-Object WorkingSet64 -Sum).Sum)/1MB)}}
# Release memory from abandoned sessions
Set-RDSessionCollectionConfiguration -CollectionName "LOBApps" `
-EnableIdleTimeout $true -IdleTimeout 30 `
-EnableDisconnectedSessionLimit $true -DisconnectedSessionLimit 10
When to add a second host
Add a session host when any one of three things is true: a single host reboot takes the whole business offline, the logon storm pushes disk queue length above 2 for more than ten minutes, or you are approaching the density figure for your size with no headroom left for a bad day. A connection broker plus a second session host is not much extra work and it converts a single point of failure into a maintenance window.
The short version: size RemoteApp like a session host, count the licences before the grace period runs out, move profiles off the host earlier than feels necessary, and treat the morning logon as your capacity test rather than your average CPU. Get those four right and RemoteApp stays predictable well past the point where desktop sessions would have fallen over.
