The Hypervisor's Cut
VPS vs dedicated performance — measuring the hypervisor's cut in practice
Virtualisation is cheap for arithmetic and dearer for I/O — and the difference between those two facts is the whole buying decision. Here is where the overhead lives and when removing it is worth a migration.
The short answer
For CPU-bound web work, a modern KVM-based VPS runs within a few percent of bare hardware, and most sites will never measure the difference. Dedicated servers earn their premium in three places: sustained disk I/O, where virtualised storage queues add latency under pressure; the complete absence of neighbours, which flattens tail latency; and full command of the hardware — every core, all memory bandwidth, the NVMe queue to yourself.
The decision metric is not traffic volume but contention symptoms: CPU steal above zero at peak, iowait climbing during heavy database work, p99 latency wandering while p50 stays put. If your monitoring shows none of those, a bigger VPS is usually the rational upgrade; if it shows them chronically, you have outgrown other people's hardware decisions.
By the HostingFast team · Reviewed 25 August 2026
~2–5%
Typical CPU cost of modern KVM virtualisation
I/O
Where the hypervisor's cut grows under load
0
Neighbours on a dedicated machine
p99
The percentile dedicated hardware flattens
The performance question between VPS and dedicated is really a question about where virtualisation's costs concentrate. Hardware-assisted CPU virtualisation is a solved problem; instructions run at effectively native speed. Storage and network I/O still traverse virtual devices and shared queues, and that is where the accounting gets interesting under load.
A dedicated server deletes the hypervisor and the neighbours in one purchase. Whether that deletion buys you measurable milliseconds — or merely a comfortable feeling — depends entirely on your workload's shape. This page gives you the numbers to decide like an engineer rather than a status shopper.
The hypervisor tax, itemised
CPU: with hardware virtualisation extensions, guest instructions execute directly on the silicon; the low-single-digit overhead lives in interrupts and context transitions. For PHP executing business logic, this line item is noise.
Storage and network: requests pass through virtio devices into queues the host schedules. Idle, the added latency is microseconds; under sustained pressure, queueing at the host layer stretches the tail. Memory: the guest's RAM is subject to the host's management, and memory bandwidth is shared with every VM on the node. None of these is dramatic alone — together they are the few percent that becomes twenty when the node is busy.
Neighbours: the variable you cannot configure away
Even a generously provisioned VPS shares silicon: cache lines, memory channels, the physical NVMe device's queue. A neighbour running a compile farm shows up in your latency histogram without ever touching your allocated cores. CPU steal measures the visible part; cache and memory-bandwidth contention are the part no guest-side metric names.
This is why dedicated hardware's real product is variance reduction. Median latency often barely moves after a migration; the p99 tightens dramatically, because the machine's only workload is yours. For a checkout flow or an API with an SLA, that tail is precisely the merchandise.
What full command of the hardware is worth
The whole NVMe queue, every memory channel, and CPU frequency behaviour that follows your load rather than the node's aggregate. Database servers notice first: buffer pools sized to real RAM with no balloon driver above them, fsync latency that stays flat through a backup, NUMA topology you can pin processes to.
There is also headroom honesty. A dedicated box's capacity is a datasheet fact; a VPS's practical ceiling depends on a host node you cannot inspect. When capacity planning matters — sales events, launches — planning against hardware you fully own is simply better arithmetic.
Deciding with your own metrics
Watch three numbers for a fortnight: %st in top (steal), %wa (iowait) during your heaviest database periods, and p99 response time against p50. Steal persistently above zero, iowait spiking with load, or a p99 that wanders while p50 holds — each points at contention a dedicated machine removes.
If all three are clean, spend the difference on a larger VPS or better caching instead; removing a hypervisor that isn't costing you anything is the most expensive upgrade in hosting. If they are dirty at your current size, the migration pays for itself in flattened tails.
Both rungs, engineered honestly
Our VPS nodes are provisioned so that steal stays where it belongs — at zero. That makes the VPS tier a place workloads can live long-term, not a waiting room for an upsell.
When your metrics genuinely argue for dedicated hardware, the migration keeps your stack intact and swaps the platform underneath: NVMe throughout, the network unchanged, and the tail latency suddenly yours alone. We would rather you upgrade on evidence than on anxiety.
- VPS nodes provisioned against steal, checked at peak not off-peak
- NVMe on both tiers — the storage class does not change, only the sharing
- Dedicated machines with the full queue, bandwidth and cores to yourself
- Migration between tiers handled by us, measured by you
Why HostingFast
Standard on every plan
Evidence-based tiering
Steal, iowait and p99 tell you which rung you need. We help you read them rather than rounding every question up to the dearest answer.
Tails you can promise
Dedicated hardware turns p99 latency from a neighbourhood property into an engineering one — the difference an SLA is made of.
The same stack either side
cPanel, LiteSpeed and NVMe on both tiers, so moving up changes the physics without changing your workflow.
Quick Start
From order to online
- 1
Instrument before you shop
Two weeks of steal, iowait and p50/p99 at real load. The decision is in that data, not in a plan comparison table.
- 2
Attribute the slowness correctly
Steal and unexplained p99 wander implicate the node; iowait under database load implicates shared storage; clean metrics implicate your stack, not your tier.
- 3
Migrate, then re-measure the tails
The success criterion for a dedicated move is a tighter p99 under identical load. If you cannot see it in a histogram, you bought comfort, not performance.
Built In
Loaded onto every plan
- CPU steal at zero on your current VPS, verified at your busiest hour
- iowait profiled during backups and heavy queries, not just at idle
- p99 tracked against p50 — divergence is the signature of contention
- NVMe confirmed on the dedicated option; a dedicated SATA box loses to a good VPS
- Capacity planned against the hardware datasheet, with real headroom arithmetic
- A rollback-capable migration plan, because tier changes should be reversible experiments
Frequently Asked
What people ask us most often
How much performance does virtualisation actually cost?
For CPU-bound work on modern KVM, low single digits — effectively invisible for PHP and web serving. The honest cost appears under sustained storage and network I/O, where virtual devices and host-side queueing stretch tail latency, and it grows with how busy the host node is. The tax is small; the neighbours are the variable.
When does a dedicated server pay for itself?
When contention is measurably costing you: chronic CPU steal, iowait during database-heavy periods, or p99 latency that misbehaves while medians look fine. It also pays where predictability itself has value — SLAs, checkout flows, capacity planning for events. Absent those, a larger VPS delivers more per unit of spend.
Is a dedicated server overkill for a WordPress site?
For most, yes — a cached WordPress site is light work, and a well-run VPS leaves nothing on the table. The exceptions are large WooCommerce catalogues with heavy uncacheable traffic, and multi-site fleets consolidated onto one machine, where the database genuinely benefits from owning the hardware.
Can a VPS have noisy neighbours even with guaranteed resources?
Yes — guarantees cover core counts and RAM, not the silicon underneath. CPU cache, memory bandwidth and the physical NVMe queue are shared, and a hard-working neighbour appears in your latency tail without touching your allocation. Steal is the visible symptom; provisioning discipline on the host is the actual cure.
Which metrics prove I have outgrown my VPS?
Three, tracked at peak: %st above zero persistently, %wa climbing with database load, and p99 diverging from p50. One bad afternoon is weather; the same pattern across two weeks is climate, and climate is what justifies a migration.
Keep reading
VPS hosting
Reserved resources on nodes provisioned so the guarantee means something.
Dedicated servers
The whole machine: every core, the full NVMe queue, a p99 of your own.
Bare metal vs virtualised
The same trade examined from the architecture side.
Shared vs VPS performance
One rung down: what reserved resources fix that sharing cannot.
Changing hosts? Run through our checklist first.
A straightforward sequence for a switch your visitors never feel: which files move first, how to shift email across without losing a single message, the right moment to repoint DNS, and the two mistakes behind almost all the downtime we get asked to rescue.
Upgrade on evidence, not anxiety
Send us your steal, iowait and p99 numbers and we will tell you honestly which tier they argue for. Migration between rungs is handled and measured.
View VPS hosting plans