CPU steal time is the time a virtual CPU sits waiting because the hypervisor did not give it physical CPU cycles. On a Linux VPS, check the st field in top. Sustained steal that lines up with slow requests can point to host-side scheduling contention. A single spike does not prove an overloaded host or a noisy neighbor.
Before you upgrade the VPS, grab a short time series and compare steal with application response times, local CPU use, memory pressure and disk activity. This guide covers how to read the metric, dodge common misdiagnoses and gather useful evidence without changing production settings.
What does CPU steal time mean on a VPS?
A VPS runs as a virtual machine under a hypervisor. The guest OS sees virtual CPUs, while the host schedules those vCPUs onto physical processors. On shared CPU plans, other VMs can compete for the same physical execution resources.
When the guest is ready to run but has to wait for execution time, supported virtualization stacks can book that delay as steal. The Linux kernel’s proc documentation describes the steal field as involuntary wait. The top manual labels st as time stolen from the VM by the hypervisor.
The name sounds dramatic, but it does not mean anyone accessed your files or compromised the VPS. Steal is a scheduling metric, not a security alert. How much of it you see also depends on the virtualization stack and kernel support, so treat the number as one clue, not a full picture of the host.
Steal time vs CPU usage, I/O wait and load average
A slow server can have several bottlenecks. Treating every high number as CPU overload sends you after the wrong fix. Use the table below to separate the signals before you change your plan.
| Signal | What it tells you | What to check next |
|---|---|---|
| High user CPU, us | Guest applications are actively executing | Busy processes, application concurrency and expensive queries |
| High system CPU, sy | The guest kernel is doing substantial work | System calls, network activity and kernel-side workload |
| High steal, st | The guest reports hypervisor-related execution wait | Timing, application impact and host scheduling with the provider |
| High I/O wait, wa | CPU accounting includes idle time associated with outstanding I/O | Disk latency and activity; I/O wait alone is not a reliable disk diagnosis |
| High load average | Runnable tasks and tasks in uninterruptible sleep are accumulating | CPU demand, blocked work, storage and the number of vCPUs |
| Swap activity and low available memory | Memory pressure may be slowing the workload | Process memory, application limits and working-set size |
These conditions can stack. A busy database can burn CPU while the hypervisor also holds the VM back. A site can also feel slow because of a remote API while local CPU metrics stay quiet. Do not stretch one useful indicator into an explanation for every outage.
Step 1: check steal time with top
SSH into the Linux VPS with your normal authorized session and run:
top
Find the CPU summary line and look for st. Let it refresh a few times. Press 1 to toggle per-CPU summaries when that is supported, then press q to exit. Per-vCPU views can expose an uneven pattern that an overall average hides.
For a bounded capture, use batch mode:
LC_ALL=C top -b -d 2 -n 16
That prints sixteen frames roughly two seconds apart. Prefer later refreshes for interval comparisons instead of treating the first screen as a full incident record. The command only observes the system; it does not tune the kernel or restart services.
The output can include process names and other operational detail. Review and redact it before sharing outside your organization. For more ways to inspect a Linux server, see our Linux system monitoring guide.
Step 2: collect a longer sample with mpstat
If mpstat is already installed, it gives a compact CPU time series. It usually ships with the sysstat package. If it is missing, install that package through your distribution’s approved process. Do not copy an untrusted installation script.
LC_ALL=C mpstat -P ALL 2 30
This asks for per-CPU reports every two seconds for thirty samples. Watch %steal, and keep the other columns. They help separate unavailable CPU time from time your own applications are using.
Take one sample during the reported slowdown and another during a quiet period. Note the server timezone and the incident window so you can line up application logs and infrastructure graphs. A measurement after the site recovers cannot show what happened earlier.
For recurring issues, lean on existing monitoring instead of opening SSH over and over. Track steal next to request latency, errors and queue length. Keep the interval short enough to catch the symptom without building a heavier monitoring setup than you need.
Step 3: understand the underlying counters
Linux exposes cumulative CPU accounting in /proc/stat. According to the proc_stat manual, the CPU fields include user, nice, system, idle, iowait, irq, softirq, steal, guest and guest_nice. Steal is the eighth numeric CPU field.
These values are counters, not ready-made percentages. A large number may only mean the machine has been up for a long time. To measure an interval, compare two readings and calculate the change in each relevant counter.
Conceptually, steal percentage is the increase in steal time divided by the increase in total accounted CPU time, multiplied by one hundred. Monitoring tools need to avoid double-counting guest time that already appears in other fields. For day-to-day troubleshooting, stick with an established tool instead of writing a parser in the middle of an incident.
Also check what your dashboard actually plots. An average across vCPUs, a maximum across cores and a long rolling average can paint different pictures of the same interval. Compare like with like before deciding one tool is wrong.
How much CPU steal time is too much?
No universal percentage proves a VPS is unusable. Brief interruptions may leave a cached site looking fine, while a latency-sensitive service can suffer from smaller recurring delays. Duration, how often it returns and the application impact all matter together.
Judge against your own baseline. Does steal rise with slow requests? Does the pattern show up across multiple samples? Does it appear when the workload is stable, or only during a traffic spike that also saturates the application?
As an illustration, repeated stretches of elevated steal that line up with request timeouts deserve a closer look. One isolated refresh with a higher reading, then normal latency, is weaker evidence. These are diagnostic examples, not fixed thresholds or proof of a contract breach.
A low reading also does not certify perfect hosting. Some environments expose limited accounting, and storage or network contention can exist without high CPU steal. Keep the user-visible symptom at the center of the investigation.
Can you identify a noisy neighbor from inside the VPS?
Usually not directly. A noisy neighbor is another tenant whose resource demand hits shared infrastructure. Guest metrics can support a contention theory, but they do not name the other customer or show that customer’s workload.
High steal can fit host scheduling pressure. It does not prove deliberate overselling, identify another VM or reveal the provider’s exact allocation policy. Only the provider has the host-side telemetry to check those questions.
For the underlying hosting model, our VPS, dedicated server and cloud comparison explains virtual versus physical isolation. Sharing hardware is normal for many VPS products. What matters is whether the service meets your workload’s requirements.
What to check before contacting the provider
Rule out your own workload changes
Look for a recent deployment, backup, compression job, crawler burst or scheduled task. Check whether slow queries or too many application workers are exhausting the server. Do not kill unfamiliar processes just because they sit near the top of a list.
If the server runs a database, review that database’s own performance data alongside host metrics. Our PostgreSQL VPS guide discusses workload planning beyond the advertised core count.
Check memory and storage separately
Low available memory, swapping and slow disk operations can look a lot like CPU trouble. Storage type alone is not a performance guarantee, but our SSD vs HDD server guide explains why latency and workload patterns matter.
Separate network delay from execution delay
If only distant users report slowness, dig into routing, connection setup and external dependencies. More CPU cannot fix geographic distance. Use our low-latency server location guide to keep that distinction clear.
What can actually reduce the problem?
If the evidence points to host-side contention, ask the provider to review scheduling and whether a host migration makes sense. Moving to another node may help, but confirm downtime, address changes and backup requirements before you accept a move.
A dedicated-vCPU plan may offer stronger execution guarantees than a shared plan, but the wording varies by vendor. Ask whether the resource means a reserved hardware thread, a core, a scheduling entitlement or a usage limit. Our server CPU guide explains why architecture and workload also matter.
Adding RAM will not directly fix missing execution time. Adding shared vCPUs is not a guaranteed cure either. Target the bottleneck you measured, not the biggest number on the upgrade menu.
Inside the guest, cut unnecessary background work and avoid stacking heavy jobs when you can. That can lower demand and improve responsiveness, but it does not give you control of the hypervisor scheduler. Keep changes small, reversible and measurable.
Build an evidence bundle, not an accusation
Prepare the incident window, the affected service, request latency or error changes, CPU samples, memory and disk observations, and a normal-period baseline. Note the number of vCPUs and whether a deployment or traffic change happened.
Ask the provider to investigate host-side scheduling against that evidence. Do not send passwords, access tokens or raw logs that hold customer information. Before a migration or other substantial change, follow the VPS backup and restore guide and confirm recovery works.
Keep a simple before-and-after record
After any approved change, repeat the same observation during a comparable traffic period. Keep the sampling interval and application metrics identical. If response times improve while steal falls, that supports the diagnosis, but it still does not name a particular neighboring tenant. If performance stays poor, go back to the evidence instead of stacking more changes at once. A small, repeatable investigation beats a screenshot of one unusually busy second.
FAQ
Does CPU steal time mean my VPS was hacked?
No. Steal is a virtualization scheduling metric, not evidence of unauthorized access. Investigate security concerns using authentication logs, integrity checks and other appropriate evidence.
Can I fix high steal time by rebooting?
A reboot is not a reliable fix for host scheduling contention. Collect evidence first, check your workload and ask the provider whether host-side investigation or migration is needed.
Does zero steal time guarantee good VPS performance?
No. Reporting varies, and a VPS can still have application, memory, storage or network bottlenecks. Measure the service’s actual latency and error rate as well.
Will dedicated vCPUs eliminate every slowdown?
No. Stronger CPU allocation can reduce scheduling contention, but performance still depends on the allocation terms, processor, application, storage and network. Validate any change with the same workload.