Management Guides Troubleshooting

What Is CPU Steal on a VPS? How to Measure It and What to Do

CPU steal is time your VPS wanted the processor and the host gave it to another customer. How to measure it with vmstat, mpstat and sar, how to read it, and how to get a provider to act.

CPU steal is the share of time your virtual machine was ready to run and the physical host gave the processor to another customer's machine instead. From inside the VPS it looks like a server that is slow for no reason: load is high, yet your own processes are not using much CPU. This article explains how to measure steal, how to read the numbers and what you can do about it.

Where steal comes from

A VPS host runs many virtual machines on a fixed number of physical cores. On shared-CPU plans, providers sell more virtual cores than the host has, on the assumption that customers will not all be busy at once. When they are, the hypervisor has to make some of them wait. The waiting time is reported to each guest as steal.

Steal is therefore a property of the host and of your neighbours. Nothing running inside your VPS causes it, with one exception: some providers throttle a VM that uses its full CPU allowance for a long period, and on some platforms that throttling is also reported as steal.

How to measure it

A quick look. In top, steal is the st value at the end of the CPU line. In vmstat it is the last column:

vmstat 1 10

Skip the first line of output, which is an average since boot.

Per core. Install the sysstat package and use mpstat, which shows %steal for each virtual CPU:

sudo apt install sysstat      # Ubuntu, Debian
sudo dnf install sysstat      # AlmaLinux, Rocky Linux
mpstat -P ALL 1 10

Over a day. A single reading proves little, because steal depends on what the neighbours are doing at that moment. Sysstat can record CPU statistics every ten minutes:

sudo systemctl enable --now sysstat
# later, show today's history
sar -u

On Debian and Ubuntu, if no history appears, set ENABLED="true" in /etc/default/sysstat and restart the service. The %steal column of sar -u then shows when during the day the problem occurs.

The raw source. All these tools read /proc/stat. On the line starting with cpu, the eighth number is steal time in clock ticks since boot.

How to read the numbers

There is no official limit, and hosting documentation gives different thresholds. As a working convention:

Sustained stealWhat it usually means
0–2%Normal on a shared-CPU plan
A few percentBusy host. Usually not noticeable for a website
Around 10% and above for long periodsThe host is overcommitted, or you are being throttled. Worth a ticket
Tens of percentThe VPS is losing a large part of its CPU. Applications will be visibly slow

Two points keep the reading honest:

  • Duration matters more than peaks. A few seconds at 30% while a neighbour starts a job is harmless. Twenty percent for six hours every evening is a real loss.
  • Steal only hurts when you need the CPU. On an idle VPS the figure can read zero even on a crowded host, because your machine is not asking for time. Measure while your workload is running.

Rule out your own workload first

High load with low steal is not a neighbour problem. Look at the other columns of vmstat:

  • High us or sy: your own processes are using the CPU.
  • High wa: the server is waiting for disk.
  • si/so above zero: it is short of memory and swapping.

Our VPS triage guide covers each of those cases.

What you can do

  1. Collect evidence. Keep sar -u output for a few days, or save vmstat 1 60 with a timestamp at several times of day.
  2. Open a ticket. Providers can move a VM to a less busy host, and many will when shown numbers.
  3. Check the plan's terms. If the fair-use policy limits sustained CPU use and your workload runs at 100%, the plan is the wrong type for the job.
  4. Move to dedicated CPU. Plans sold with dedicated cores (often called dedicated vCPU, VDS or CPU-optimised) reserve physical cores for you. They cost more because nobody else can be sold the same core.
  5. Change provider if steal stays high after a migration and the provider has no other answer.

Resizing to more shared cores on the same host rarely helps: more virtual cores on a crowded machine means more cores waiting.

A ticket that gets results

Subject: Sustained CPU steal on VPS [name / IP]

My VPS has shown CPU steal between X% and Y% daily from HH:MM to HH:MM (UTC)
since [date]. Output of "sar -u" for the last three days is attached.
My own CPU use in the same periods was around Z% (user + system), and
I have ruled out memory and disk as causes.

Could you check the host node and migrate the VPS to a less loaded one?
Please also confirm whether this plan applies CPU throttling.

Fill in real figures. A request with a time range and measurements is hard to answer with a template.

Frequently asked questions

Can I fix CPU steal myself?

No. It is decided by the host. You can reduce how much CPU you need, which makes steal matter less, or move to another host or plan.

Is steal the same as throttling?

They are different mechanisms that can look alike. Steal from contention comes and goes with the neighbours. Throttling appears after your own heavy use and continues while that use lasts. Ask the provider which applies.

Does steal affect dedicated servers?

No. A dedicated server has no hypervisor sharing its cores, so the value is always zero.

Why does my benchmark score vary so much between runs?

On shared CPU that is steal showing up in a different form. Run benchmarks at several times of day and record %steal alongside them.

Share story:

Continue reading other guides on How VPS.