Bare Metal vs VM: Are VMs slower than Bare Metal?

Bare metal vs cloud
19 October 2020

Bare metal servers typically provide more powerful and consistent performance to virtual machines.

At its most basic, server performance is how much you can get out of your computer for the tasks you need doing. Various factors, from both hardware and software, can impact the performance of a server. Lower performance can cause a reduction in productivity or bad user experience. For example, a poorly performing game server can translate into latency, experienced by the gamer as hiccups or lag in their game.   

Table of Contents

Bare metal and VM explained

What is bare metal?

Bare metal server refers to the literal physical machine, also known as a dedicated bare metal server. A bare metal server is dedicated to a single client known as a “tenant”. Bare metal can process more data than any other solution precisely because you can allocate all of the server’s resources to your dynamic workload. To understand when you should use bare metal servers, check out our guide on where bare metal hosting makes sense.

What is a VM?

A virtual machine can be created inside a bare metal server. A virtual machine forms a new software layer within the server which the tenant can use as an independent computer. Splitting physical machines into VMs is called virtualization. A single physical machine can therefore run multiple VMs. However, the VMs all share the same physical resources, such as RAM and CPU.

Performance of VMs

While VMs offer businesses a lot of benefits through their flexibility, they are often prone to latency due to two main factors:

Hypervisor Tax

The virtualization technology that creates the virtual layer is called a hypervisor – a layer of software that manages and allocates the physical resources between the VMs. Running a hypervisor comes with a “tax” on performance. Simply installing and running a hypervisor automatically creates an overhead of between 5 – 10% on the physical server resources before any other usage has taken place.

Noisy Neighbors

Once a hypervisor is installed, virtual machines can share the physical resources of a bare metal server. If a neighboring tenant runs a resource-heavy application, this can impact the performance of your VM as more of the communal resources are redirected. You can expect an additional 20-30% loss of performance due to the noisy neighbor effect. The variability of other tenant’s usage is also what contributes to the inconsistency of VMs.

Server consistency & containers

Performance is only meaningful if it is consistent. During a given period of time, a VM may be able to perform more than a bare metal server at its peaks, but average out lower overall. Bare metal can provide consistent performance because the resources don’t have any other variable input from other tenants.

Containers

Containers – a method of running and transporting applications safely between different IT environments – can be installed directly on a physical server, which has been found to significantly improve containerized workloads. Container performance on bare metal can be 25-30% better. A study by Centurylink (now known as Lumen) found that popular container orchestration system Kubernetes had 3x lower latency on bare metal compared to a VM.

VMs can offer a reduced performance to bare metal due to the overhead on performance from the hypervisor installation, and the potential drain on physical resources from neighboring VMs. Consistent results can enable businesses to forecast trends, player numbers and other data to effectively budget and plan for the future. Workloads that require high performance and consistency can benefit from bare metal servers.

Hybrid model

Dynamic or variable workloads could also investigate a hybrid model solution, like i3D.net’s bare metal cloud, which offers a mix of bare metal base and cloud burst-ability.

Main Take-Aways

Various factors, from both hardware and software, can impact the performance of a server. While VMs offer businesses a lot of benefits through their flexibility, they are often prone to latency due to two main factors: the “hypervisor tax” and the “noisy neighbor” effect.