Cover: Bare Metal vs VM: What the Overhead Benchmarks Show
October 20, 2026·13 min read·2,905 words·

Bare Metal vs VM: What the Overhead Benchmarks Show

AWS measured VMs within 1% of metal. IBM saw over 40% on MySQL. StorageReview averaged 61% to 92%. Why bare metal vs VM overhead depends on workload.

No single overhead figure exists for a virtual machine. AWS measured Nitro instances within 1% of its own metal instances on four HPC codes. IBM Research found KVM added more than 40% overhead on a MySQL test. StorageReview's 2024 hypervisor comparison averaged between 61.58% and 92.34% of bare metal, depending on the hypervisor and its setup. So the useful question is not which number is right, but what each test measured, why the numbers diverge, and how to choose between a dedicated server and a VM for your workload. A decision table sits near the end, alongside a checklist for testing your own stack.

The short answer: two definitions and one rule

A bare metal server runs your operating system directly on physical hardware as the sole tenant. A virtual machine runs a guest operating system on a hypervisor, which divides one host among multiple guests. A cloud instance or a VPS is a VM in this sense, even when the provider avoids the word.

Overhead depends on the workload and hypervisor tuning, not the label. Some workloads lose almost nothing inside a VM. Others take a heavy hit. And telling them apart means looking at the execution path, not stopping at the packaging.

Type 1 vs type 2, and why 'bare metal hypervisor' is a different thing

The terminology trips people up, which helps explain why searches for 'bare metal hypervisor' and 'bare metal vs hypervisor' become so muddled.

IBM's hypervisor explainer defines a type 1 hypervisor as software running directly on the physical hardware, also referred to as a bare-metal or native hypervisor. A type 2 hypervisor runs as an application inside an operating system. That is the entire distinction, and it matters.

A 'bare metal hypervisor' sits directly on the hardware to host VMs. A 'bare metal server' has no hypervisor between your kernel and the hardware. The same two words produce opposite realities for your workload. So anyone running on a bare metal hypervisor is still running VMs.

StorageReview describes KVM as turning Linux into a type 1 hypervisor by building virtualisation directly into the kernel. Its 2024 comparison also evaluated VMware ESXi 8, Microsoft Hyper-V, and Proxmox (which wraps KVM). The IBM Research paper expected ESX, Hyper-V and Xen to match KVM closely because all rely on the same hardware acceleration features. But that was an expectation from 2015, not a measurement, and the 2024 figures show how much implementation detail can change the result.

What the benchmarks actually measured

Three sources dominate these discussions. Each measured different workloads on different hardware across different years, so align their contexts before comparing numbers. Otherwise, the table flatters certainty it has not earned.

Source and datePlatformWhat was comparedFinding
AWS HPC Blog, 2021c5n.18xlarge (Nitro hypervisor) vs c5n.metal, same hardwareWRF, OpenFOAM, GROMACS, High Performance Linpack at 16 instances (576 cores)Within 1% in all evaluated cases; each point is the average of 3 runs
IBM Research (Felter et al.), ISPASS 2015IBM System x3650 M4, two Xeon E5-2665, 256 GB RAM, Ubuntu 13.10, QEMU 1.5.0, KVM guests with 32 vCPUsKVM and Docker vs native, including SysBench oltp on MySQL 5.5.37CPU and memory overhead almost none; KVM overhead above 40% on MySQL in every measured case
StorageReview, 13 June 2024Dell R760, bare metal baseline, one VM given the whole host, default settingsKernel compile, Apache, OpenSSL, SQLite, Stream, FIO, 7-Zip via Phoronix Test SuiteAverages vs bare metal: Hyper-V 92.34%, ESXi 89.36%, Optimized Proxmox 85.16%, KVM on RHEL 79.55%, stock Proxmox 61.58%

Two caveats apply. The AWS figures come from AWS testing its own platform, using an in-house hypervisor tuned to minimise overhead on HPC codes. Meanwhile, the IBM paper relied on 2013-era hardware and software, so treat its raw numbers as historical data from that generation rather than a baseline for modern silicon.

StorageReview also ran a test where VMs received a standard resource allocation rather than the full host, normalising results against the top performer in each category. The averages shifted: ESXi 96.4%, Optimized Proxmox 81.7%, KVM on RHEL 79.79%, Hyper-V 63.27% and stock Proxmox 59.69%. The order changed completely. That shift alone should stop anyone treating a single table as definitive.

Some individual results from the all-resources run are worth seeing, because the averages hide them:

  • FIO 4k random read/write: ESXi 57.41% and 55.27%, Hyper-V 72.95% and 85.71%, KVM on RHEL 74.60% and 85.37%, stock Proxmox 54.71% and 44.71%, Optimized Proxmox 98.57% and 91.49% of bare metal.
  • OpenSSL: ESXi, Hyper-V and KVM on RHEL all landed close to bare metal, stock Proxmox managed 5.33%, and Optimized Proxmox 98.91%.
  • Stream memory bandwidth: ESXi 98.30% and Hyper-V 99.01%, but KVM on RHEL 74.60%, stock Proxmox 76.24% and Optimized Proxmox 71.04%.
  • Apache: ESXi 113.64% and Hyper-V 129.62%, which StorageReview put down to hardware accelerators in newer chips.

A VM actually beat bare metal here. Keep that in mind: the benchmark reflects specific configuration choices and hardware features as much as it measures any virtualisation tax.

StorageReview adds another caveat: the initial run and subsequent rerun resist direct comparison because the host changed from a Sapphire Rapids Xeon 6430 to an Emerald Rapids Xeon 8580. The testers also could not account for their own SQLite results, where stock Proxmox at 85.27% outperformed Optimized Proxmox at 68.86%. Real-world benchmarks are messy. Pretending they produce clean constants helps no one.

Why the numbers disagree

Strip away the branding and the disagreement comes down to four mechanical things.

Every I/O operation pays a toll

IBM offers the clearest explanation in the literature. Pure CPU and memory instructions execute near native speeds, meaning VMs 'impose almost no overhead on CPU and memory usage'. The performance penalty hits I/O and OS interactions instead, showing up as extra instruction cycles per I/O call. Small I/O operations suffer significantly more than large ones, since large transfers amortise the per-operation penalty. This inflates I/O latency and burns CPU cycles that would otherwise run application code.

This explains why a database or a 4k random-I/O test degrades while a compute-bound simulation runs flat out. IBM notes that production software rarely batches operations into large blocks. But there is an important detail: the MySQL test instance had enough RAM to avoid physical disk access entirely. The 40%-plus penalty came purely from the I/O and OS-interaction path, including network calls, IPC, and filesystem overhead, rather than storage media latency.

Hardware offload changes the equation

AWS states that Nitro cards handle VPC networking, EBS access, and local NVMe storage, removing the need to spend host CPU cycles on those tasks. With I/O handled on dedicated cards, the Nitro hypervisor remains lean and quiescent, activating only when an instance explicitly requests work.

A within-1% result on Nitro demonstrates the efficiency of offloaded I/O hardware. It says little about how a standard VM behaves on commodity hardware. Change the I/O path, and the performance profile changes with it.

Tuning matters a lot

StorageReview's data confirms this disparity. Its Optimized Proxmox configuration used the host CPU model, enabled NUMA, selected the q35 machine type, booted OVMF (UEFI) firmware, and turned on Write Back caching with SSD emulation. The competing hypervisors received only baseline resource allocations. Consequently, stock Proxmox averaged 61.58% while Optimized Proxmox hit 85.16% on identical software.

IBM reached the same conclusion in 2015: solid KVM CPU performance demands careful configuration of large pages, CPU model, vCPU pinning, and cache topology. Those settings were poorly documented and required trial and error. If your VM provider configured them for you, fine. If you run your own KVM host with defaults, your benchmark simply measures those defaults.

Resource allocation changes the ranking

Allocating an entire physical node to one VM, as StorageReview did initially, behaves differently from partitioning hardware among concurrent guests. Hyper-V posted 92.34% in the single-tenant test, but dropped to 63.27% under shared conditions. If your production deployment partitions hardware, the dedicated-host benchmarks offer no guidance.

Noisy neighbours and CPU steal time

Clean synthetic tests ignore an inevitable operational cost: noisy neighbours. When running on shared hardware, the hypervisor arbitrates CPU time, and your workload will periodically wait. That waiting is part of the platform, not an edge case.

The Linux proc_stat man page defines the steal metric, present since Linux 2.6.11, as stolen time: cycles consumed by other operating systems in a virtualised environment. Red Hat's documentation explains it from the guest perspective: cycles the guest requested but could not obtain because the hypervisor scheduled execution time for another tenant.

How to read it

Steal is reported in the CPU time fields of /proc/stat (it is the eighth value on the cpu line, after the label). Utilities like top and vmstat show it for you, as %st in top and the st column in vmstat. Red Hat notes it cannot be switched off, so it is always being counted.

A quick way to look:

  1. Run top and read the %st figure in the CPU summary line.
  2. Or run vmstat 5 and watch the st column over several samples.
  3. Check it during your busy period, not at 3am when nothing is happening.

What a sustained non-zero value tells you

Red Hat says large amounts of steal indicate CPU contention, which can reduce guest performance. Its suggested remedies are raising the guest's CPU priority or quota, or running fewer guests on the host. Neither is something you control on a typical VPS. A 2015 Linux kernel patch cover letter frames it the same way: customers can use steal as an indicator that their guests are being throttled, and providers overcommit CPU.

A blip means nothing. A steady, visible figure while your own processes are busy means the host is oversold, or your plan's CPU share is smaller than you assumed. That is the point where moving to a larger plan, a dedicated-core plan, or a bare metal server stops being a luxury. And on a bare metal server there are no other guests to lose time to, so the field has nothing to report.

Provisioning and agility: where VMs win outright

Benchmarks reward bare metal. Calendars reward VMs.

AWS lists faster provisioning times among the benefits of virtualised hardware, and OVHcloud's bare metal vs VM explainer makes the same case: VMs are more flexible and scalable, and you can scale resources up or down as needed. Even AWS's own bare metal instances come with a caveat. Its EC2 documentation says the server goes through a boot process that verifies all hardware and firmware, so it can take up to 20 minutes or more from the time the instance enters the running state until it is available over the network.

That gap matters if you create and destroy environments all day. A CI runner, a staging copy for a client, a throwaway test box: all of these are natural VM jobs, because nobody wants to wait around for hardware checks.

The other agility advantages are structural. A VM is a file-backed object on a hypervisor, so snapshotting it, cloning it and moving it between hosts are things the platform can do for you. StorageReview notes Proxmox offers advanced backups and snapshots, and the kernel patch cover letter above mentions migrating guests off a loaded host as something providers can do. A physical server has none of that built in. You rebuild from backups and configuration management, and that takes longer.

When bare metal is the right call

Here is where I'd actually reach for a dedicated server.

  • Licensing or support rules that forbid virtualisation. AWS says metal instances are most commonly used for workloads that cannot run virtualised for licensing reasons, and its docs mention applications that need a non-virtualised environment for licensing or support. If the vendor's terms say no VMs, the benchmark debate is over.
  • Hardware features that need direct access. The same AWS docs point to low-level features such as Intel VT that are not available or fully supported in virtualised environments. If your software needs to run its own hypervisor or touch hardware directly, you need the hardware.
  • High-rate small I/O and latency-sensitive databases. IBM concluded that KVM adds some overhead to every I/O operation, significant for small I/Os and negligible once amortised over large ones, which makes it less suitable for latency-sensitive workloads or those with high I/O rates. Its MySQL result is the sharpest example, with the 2013-era caveat attached.
  • Sustained full-host utilisation. If one workload will use nearly every core all day, the sharing that makes VMs cheap no longer helps you, and you are only paying for a hypervisor layer you don't use. (This one is my reasoning rather than a measurement from the sources, so check it against your own numbers.)

Going the other way matters just as much. If your workload is bursty, small, or changes shape often, a VM is usually the better fit, and HostList's VPS hosting rankings show where that makes sense. If you are weighing a VPS against a whole server rather than a hypervisor against no hypervisor, VPS vs dedicated server works through that choice by workload, and the dedicated hosting rankings list providers for the cases where metal wins.

The workload decision table

Every cell below points back to a cited measurement or documented behaviour from above. Read the 'why' column as the evidence, not as a rule.

WorkloadBare metalVMWhy
CPU-bound computeFine, but little to gain on a well-built hypervisorOften close to metal. AWS saw within 1% on four HPC codes (Nitro, 2021)IBM (2015) found negligible CPU overhead, but StorageReview (2024) saw stock Proxmox at 5.33% on OpenSSL, so check the defaults
Memory bandwidthBaselineStream ranged from 71.04% to 99.01% of bare metal depending on hypervisor (StorageReview, 2024)IBM (2015) found negligible memory overhead on its older hardware; configuration and generation change the answer
Small random disk I/OPreferred when the I/O rate is highFIO 4k random read/write ranged from 54.71% and 44.71% (stock Proxmox) to 98.57% and 91.49% (Optimized Proxmox)IBM (2015): extra cycles on every I/O operation, so small I/Os suffer most
DatabasesPreferred for latency-sensitive or high-rate workloadsIBM saw KVM above 40% overhead on MySQL (2013-era hardware)IBM (2015) calls KVM less suitable for latency-sensitive or high-I/O-rate work. Benchmark your own engine and storage
Latency-sensitive networkPreferred when every microsecond countsWorkable when I/O is offloaded: AWS Nitro cards handle network, EBS and local NVMeIBM (2015) reports line-rate 10 Gbps on a single core under KVM with 2013-era kit, but flags latency as the weak point
Licensing-bound or hardware-feature softwareRequired when the vendor forbids virtualisationNot an option in those casesAWS documents licensing and support needs, plus features like Intel VT, as the reasons for metal
Dev and testSlow to create and resetPreferred: faster provisioning (AWS), flexibility and scaling (OVHcloud), snapshots and clonesAWS docs: a bare metal instance can take up to 20 minutes or more after running state before it is reachable
Bursty web appsPays for idle capacityPreferred, as long as steal time stays near zeroRed Hat: sustained steal means CPU contention. Scaling and provisioning favour VMs (OVHcloud)

How to test your own workload

Published figures narrow the search. They don't replace a test on your own stack.

  1. Pick a benchmark that looks like your real traffic or job, not a synthetic one that stresses something you never use.
  2. Run the identical test on both a VM and a bare metal machine, same operating system, same software versions, several runs each.
  3. During the VM run, watch %st in top or the st column in vmstat, and record disk and network latency, not just throughput.
  4. Run the test at the concurrency you expect in production, since IBM's MySQL gap showed up across all its concurrency levels and StorageReview's ranking changed between allocation setups.
  5. Divide cost by units of work done (requests, builds, transactions), not by hours of server time. A cheaper box that does less may cost more per unit.
  6. Retest after any hypervisor, kernel or hardware-generation change. These results age quickly.

Frequently Asked Questions

Why did some VMs score above 100% of bare metal?

In StorageReview's Apache test, ESXi reached 113.64% and Hyper-V 129.62% of bare metal. StorageReview attributed this to hardware accelerators in newer chips. It's a reminder that these are comparisons between specific software and hardware setups, not pure measurements of a virtualisation tax.

Do containers change this picture?

Only slightly for this decision. The IBM paper compared KVM with Docker as well, and found Docker's difference from native approached 2% at higher MySQL concurrency, against KVM's 40%-plus, but that is one 2013-era test and not a general ruling on containers.

Can I plan capacity with a standard overhead percentage?

Not reliably. The measured range across the sources runs from within 1% to above 40% on a single workload, and StorageReview's own averages moved with tuning and allocation. A rule-of-thumb percentage will be wrong for somebody. Benchmark your workload.

Does a type 1 hypervisor remove the overhead?

No. A type 1 hypervisor runs directly on the hardware, per IBM, but it is still a hypervisor with guests. StorageReview's tests used type 1 style hypervisors throughout and still measured gaps against bare metal.

The sharpest caveat from the sources is that the same label, 'VM', covered results from within 1% to above 40%, so the label tells you nothing until you know the I/O path, the tuning and the workload.

HostList on LinkedIn
More independent hosting data

Follow HostList for new rankings, original research, and changes across the hosting industry.

Gautam Khorana
Gautam Khorana
Founder, HostList.io

Over 10,000 websites launched. Years managing thousands of client sites. Built HostList because the world deserves honest hosting advice.

LinkedIn →

MENTIONED HOSTS

RELATED ARTICLES

RadixSponsorPremium names that work like prime real estate400,000+ short, memorable premium domains across .tech, .store, .online, .site and more. 20,000+ already sold.See Radix premiums
.tech domains from RadixSponsor.tech: the address for what you buildPremium .tech names like cloud.tech and micro.tech, from Radix. Short, dictionary-word domains for tech brands.See premium .tech
.host domains from RadixSponsor.host: a domain that says what you doPremium .host names for hosting companies and infrastructure brands, from the Radix registry.See premium .host
GreenGeeksSponsorEco-Friendly WordPress Hosting DealFast WordPress performance backed by expert 24/7 support, free migration, daily backups and built-in security.See GreenGeeks
HostPapaSponsorFast, Reliable, & Affordable Web HostingLaunch, grow and manage your website with reliable hosting, easy tools and 24/7 PapaSquad support.See HostPapa

Promoted placement. Does not affect HRI, ranking order or eligibility.