Performance Monitor ships with every Windows Server licence and writes binary counter data you can replay, export, and alert on — no agent, no subscription. The difficulty is not turning it on; it is choosing a counter set small enough to read under load and large enough to tell a CPU problem from a disk problem. This guide gives a set that fits on one screen, the built-in data collectors worth keeping, and the alerting that turns counters into action.
The Minimum Counter Set Worth Watching
Twenty counters is plenty. More than that and you will stop looking. These are grouped so each pair answers one question:
| Question | Counter | Bad If |
|---|---|---|
| Is the CPU saturated? | \Processor(_Total)\% Processor Time | Sustained > 80% |
| Is it queued? | \System\Processor Queue Length | > 2 × vCPU count sustained |
| Is the cache thrashing? | \Memory\Available MBytes | < 200 MB for long periods |
| Is the disk the limit? | \PhysicalDisk(_Total)\Avg. Disk sec/Transfer | > 0.020 s (20 ms) |
| Is the disk queueing? | \PhysicalDisk(_Total)\Avg. Disk Queue Length | > 2 × virtual disks |
| Are we paging? | \Memory\Pages/sec | > 50 sustained |
| Is the network saturated? | \Network Interface(*)\Bytes Total/sec | Near rated link speed |
Disk latency is the counter people skip and then regret. On a Windows VM, a 20 ms average transfer time is a real ceiling: above it, IIS requests queue, .NET thread pools starve, and everything looks like “the app is slow”. A 40 ms average on a disk advertised as NVMe is a provider-side problem, not an application one.
Collecting a Baseline From PowerShell
Capture a 10-minute baseline into a .blg file you can keep. The file is small and opens on any Windows machine, including your own laptop:
$counters = @(
'\Processor(_Total)\% Processor Time',
'\System\Processor Queue Length',
'\Memory\Available MBytes',
'\Memory\Pages/sec',
'\PhysicalDisk(_Total)\Avg. Disk sec/Transfer',
'\PhysicalDisk(_Total)\Avg. Disk Queue Length',
'\PhysicalDisk(_Total)\Disk Bytes/sec',
'\Network Interface(*)\Bytes Total/sec'
)
Get-Counter -Counter $counters -SampleInterval 5 -MaxSamples 120 |
Export-Counter -Path C:\perf\baseline-$(Get-Date -Format yyyyMMdd-HHmm).blg -Force
Take two baselines: one at idle and one under the workload you care about. The delta between them tells you which resource moves first. That ordering — not absolute values — is what lets you add vCPU, RAM, or disk without guessing.
Data Collector Sets Worth Keeping
Windows ships three Data Collector Sets under Data Collector Sets > System: System Diagnostics, System Performance, and Wireless Diagnostics. The first two are worth keeping. System Performance runs for 60 seconds on demand and produces an HTML report that flags the top resource consumers — the fastest way to answer “what was using CPU at 3 a.m.” without a permanent collector:
logman start "System\System Performance" -ets
Start-Sleep -Seconds 65
logman stop "System\System Performance" -ets
Get-ChildItem 'C:\PerfLogs\Admin\System Performance\*\*.html' |
Sort LastWriteTime -Desc | Select -First 1
For long-running collection, create a user-defined collector that rolls on size and age — a forgotten collector filling the C: drive is a classic reason a VPS goes read-only at 3 a.m.:
logman create counter vps-baseline -f bin -o C:\perf\vps-baseline `
-si 15 -max 300 -cnf 7d `
-c "\Processor(_Total)\% Processor Time" "\Memory\Available MBytes" `
"\PhysicalDisk(_Total)\Avg. Disk sec/Transfer"
Turning Counters Into Alerts You Will Read
Performance Monitor alerts fire a data collector when a counter crosses a threshold. Crude, free, and genuinely useful for three things, all written to the Application log your log shipping already watches:
- Avg. Disk sec/Transfer > 0.050 s for 5 consecutive samples — sustained disk latency.
- Available MBytes < 256 for 10 minutes — memory pressure or a leak.
- LogicalDisk(C:)\% Free Space < 12 — early warning before update staging and logs hit zero.
Reading a .blg File Without the GUI
For quick triage, pull summary statistics straight out of the file with Import-Counter instead of clicking through charts:
Import-Counter C:\perf\baseline-*.blg |
ForEach-Object { $_.CounterSamples } |
Group-Object Path |
Select-Object Name,
@{n='Min';e={[math]::Round(($_.Group | Measure-Object CookedValue -Minimum).Minimum,3)}},
@{n='Avg';e={[math]::Round(($_.Group | Measure-Object CookedValue -Average).Average,3)}},
@{n='Max';e={[math]::Round(($_.Group | Measure-Object CookedValue -Maximum).Maximum,3)}} |
Format-Table -AutoSize
The average-with-max view usually identifies the culprit in one read. If disk latency averages 3 ms but spikes to 400 ms every ten minutes, you are not disk-bound in the aggregate — you are hitting a snapshot, a backup, or a provider-side storage pause, and more vCPU will not help.
Where PerfMon Lies on a VPS
Virtualisation changes what some counters mean. Inside a Hyper-V or KVM guest, \Processor Information(_Total)\% Processor Time reflects the guest’s view and may not show steal time; the host can throttle you while the guest never looks busy. The signal to watch is a processor queue length that grows while processor time sits near 60% — the signature of CPU limits or noisy neighbours. Likewise “Avg. Disk Queue Length” is measured at the virtual disk, so a value of 4 might mean one physical device serving four tenants.
Compare against the provider’s own metrics tab. If the guest reports low utilisation while the host reports sustained steal, the fix is a plan change, not a service restart. The Windows VPS hosting plans and specifications map vCPU and RAM to IIS, SQL Server, and RDS workloads so you can tell whether the answer is a bigger plan or a better configuration.
A One-Page Monitoring Routine
Set the vps-baseline collector to start at boot, sample every 15 seconds, roll weekly, and keep 300 MB. Review the .blg after any complaint. Keep the three alerts enabled. That is the whole routine — one collector, three alerts, one weekly review. Anything more elaborate for a single VPS is usually a monitoring product looking for a justification, and the Windows VPS hosting plans and specifications explain which tiers include the metrics dashboards that pair well with these counters.



