10 min read

Learning Objectives

By the end of this module, you will be able to:

  • Choose a compute class per tier along four axes: core isolation, CPU-family control, Block Storage attachment, and operational model
  • Evaluate the fixed template of Cubes and the hardware limits of Confidential VM as design inputs, not as defects
  • Plan a change of compute class or CPU family as a migration or a restart, not as a setting
  • Configure FinCorp's Dedicated Core Servers in the Data Center Designer (DCD) with CPU family, cores, RAM and SSH credentials

Unit 4.1: Compute Class Selection

Introduction

Unit 3.1 placed four Dedicated Core Server shells on FinCorp's application and data LANs and left their sizing open. This unit makes the decision those shells were waiting for: which compute class each tier runs on, and with which CPU family and size.

The decision matters because a compute class is not a toggle. Compute Engine types are fundamental architectural choices in the DCD and cannot be toggled directly; moving between them means moving storage to a new instance. Choose per tier now, and the later units on disks, scaling and migration build on a stable base. The unit ends by sizing FinCorp's servers in the DCD.

1. The Four-Way Decision

1.1 The Classes Side by Side

Four general-purpose classes are available for VDC workloads. The Cloud GPU VM is a fifth class with its own templates; Unit 6.5 treats it. The following table compares the four on the axes that separate them.

Axis Dedicated Core Server vCPU Server Cubes Confidential VM
Core isolation Exclusive CPU use Shared CPU use Shared cores Dedicated CPU cores, memory encrypted with AMD SEV-SNP
CPU family Selected at provisioning; changeable with a restart Not selectable, not changeable Not selectable Derived from the uploaded image; read-only
Size 1 to 62 cores, 0.25 to 230 GB RAM 1 to 60 vCPUs, 0.25 to 230 GB RAM Fixed template, 1 to 16 vCPUs, 2 to 64 GB RAM vCPU count and RAM fixed at creation
Block Storage Up to 24 PCI connectors, shared by NICs and storage Up to 24 PCI connectors, shared by NICs and storage One NVMe volume plus up to 23 add-on Block Storage devices Startup volume from the confidential image
Uptime SLA 99.95% 99.95% 99.9% No live migration; host maintenance causes downtime

As shown above, the classes differ in four ways that each map to a tier requirement. Additional RAM sizes beyond 230 GB are available on request for Dedicated Core and vCPU Servers.

1.2 Core Isolation

On a Dedicated Core Server, IONOS CLOUD maintains dedicated resources for each customer, and you do not share your physical CPUs with other clients. On a vCPU Server, the physical server is shared. Unit 2.4 priced the difference; the architectural consequence is predictability. A tier whose latency is part of a service commitment, or whose CPU consumption is steady and high, belongs on exclusive cores. A tier that tolerates contention, such as development, test or a lightly used internal tool, belongs on shared cores.

One detail affects capacity planning. On most CPU families, a single physical core is exposed to the guest as two logical cores. The Intel Sierra Forest family uses high-throughput cores without Hyper-Threading. Size a CPU-bound workload against the family it runs on, not against a core count alone.

1.3 CPU-Family Control

Explicit CPU architecture selection is only available when deploying Dedicated Core Servers. The available families are region-dependent. In de/fra, where FinCorp's production VDC sits, they are AMD EPYC Milan and the Intel families Haswell/Broadwell, Skylake and Ice Lake; de/fra/2 offers AMD EPYC Turin and Intel Sierra Forest. Intel Sierra Forest and AMD EPYC Turin require UEFI for booting, which Unit 4.2 turns into an image requirement.

You can change the CPU family of a running Dedicated Core Server, a capability named Core Technology Choice, but you have to restart the server first. Treat a family change as a planned operation with a maintenance window. On a vCPU Server, there is no such lever: the family cannot be chosen at creation or changed later.

1.4 Block Storage Attachment

Dedicated Core and vCPU Servers have up to 24 PCI connectors, and NICs and storage volumes share them. A server with four NICs therefore has room for 20 volumes. Count both when you design a server that needs many network legs and many disks.

Cubes take a different shape. Each Cube comes with one direct-attached NVMe volume, sized by the template, on the same physical host. That volume is set upon provisioning and cannot be unmounted or deleted from the instance. It is protected by a software RAID system, not by the replicated Block Storage described in Unit 4.2. A Cube can additionally attach up to 23 Essential, Balanced or Performance Block Storage devices, for 24 devices in total, and those added devices can be unmounted and deleted at any time. A Cube is therefore not a storage dead-end. Its trap is the template.

1.5 Operational Model and SLA

The template of a Cube, its vCPU, RAM and direct-attached storage size, cannot be changed after the Cube is provisioned. Live vertical scaling on a Cube covers NICs and Block Storage only. Growth means a new Cube from a larger template, which is a migration. Cubes also carry a lower uptime commitment (99.9%) than Compute Engine (99.95%), a suspended Cube continues to incur costs, and Cubes fall outside the scope of the IT-Grundschutz certificate (Unit 1.4). The Cubes documentation names development and testing environments, website hosting and simple web applications as the fitting workloads.

2. Confidential VM as a Compute Class

Unit 1.4 chose Confidential VM as the control for data in use. As a compute class, it imposes limits that come from the AMD SEV-SNP hardware and cannot be changed:

  • Location: available in de/fra/2, on AMD EPYC Turin hosts.
  • No live migration: host maintenance triggers user-visible downtime, and complete re-attestation is required after the host restarts.
  • Fixed shape: you cannot add or remove vCPUs after creation, and RAM is set at provisioning and cannot be resized. A different vCPU count requires a new VM with an updated image.
  • No hot-plug and no suspend: adding or removing RAM, NICs or disks requires a stop, a change, a restart and a new attestation.
  • No in-guest restart: running reboot inside the VM terminates it; restart through the DCD Reset action or the API reboot endpoint.
  • No cloud-init, no startup-volume snapshots: embed configuration in the image at build time, and keep your image and build pipeline outside IONOS CLOUD.

The rate is 0.047 € per core and 0.005 € per GB of RAM per hour. Design the tier to tolerate restarts: run at least two instances behind a balancer, and treat the image pipeline as the source of truth, because there is no snapshot to fall back on.

3. Changing Class Is a Migration

Because types cannot be toggled, a wrong class choice is corrected by moving storage. The following table lists the supported paths.

From To Method 1: Storage re-attachment Method 2 (snapshot) Method 3: IONOS CLOUD Backup Service (Acronis)
vCPU Server Dedicated Core Server Supported* Supported* Supported
Dedicated Core Server vCPU Server Supported* Supported* Supported
Cube vCPU Server or Dedicated Core Server Secondary volumes only (boot volume locked) Supported* Supported
vCPU Server or Dedicated Core Server Cube Secondary volumes only (boot volume locked) Supported (check size limits)* Supported

As shown above, leaving a Cube always involves a snapshot or a backup for the boot volume. The entries marked with * (Methods 1 and 2) also require the source and target CPU to support the same boot firmware (BIOS or UEFI); on a mismatch, use the Backup Service. Every path is a new instance, so plan new addresses, a cutover and a rollback.

FinCorp applies the classes per tier:

Tier Class Why
Portal servers on fc-app Dedicated Core Server Within the IT-Grundschutz scope; predictable latency; covered by Cloud Savings Plans (Unit 2.4)
Data-tier servers on fc-data Dedicated Core Server, AMD EPYC Milan Exclusive cores; a fixed CPU family for the licensed database software
Fraud scoring, in a VDC in de/fra/2 Confidential VM Data in use must stay opaque to the operator (Unit 1.4)
Development and test in the non-production contract vCPU Servers; Cubes for short-lived build agents Cost; no compliance scope requirement

DCD Implementation Walkthrough

You will size the four Dedicated Core Server shells that Unit 3.1 placed in fc-prod-fra-core. This realises the first two rows of FinCorp's class table. Disks, images and cloud-init follow in Unit 4.2.

Prerequisites: the servers fc-app-01, fc-app-02, fc-data-01 and fc-data-02 from Unit 3.1, their tier NSGs fc-nsg-app and fc-nsg-data from Unit 3.2, and a public SSH key stored in the SSH Key Manager.

Build goal: Provision a dedicated-core server shell (class, cores/RAM, credentials).

Steps (in the Data Center Designer):

  1. Open fc-prod-fra-core and select fc-data-01 in the Workspace. The Inspector pane shows its Settings tab.
  2. Keep the Name unique within the VDC. Keep Availability Zone at Zone 1; for fc-data-02, Zone 2. The server zone can be changed after provisioning; the storage zone in Unit 4.2 cannot.
  3. In CPU Architecture, choose the family for the tier from those offered in the location: AMD EPYC Milan for the data tier. Record it in the architecture decision record, because a later change needs a restart.
  4. Set Cores within the class limit of 62. Leave headroom below the limit, because Unit 4.3 grows this tier live.
  5. Set RAM in steps of 0.25 GB. For a Windows server, start at 8 GB: servers created from IONOS CLOUD Windows images with more than 8 GB may not start, and you scale up after the first boot.
  6. Check Network Security Groups: the tier group from Unit 3.2 stays assigned.
  7. In SSH Keys, select the premade key from the SSH Key Manager, or paste the public part of a key as an Ad-hoc Key.
  8. Repeat steps 1 to 7 for fc-data-02, then for fc-app-01 and fc-app-02 with the portal sizing.
  9. Select PROVISION CHANGES, confirm with your password and select Provision Now.

Common mistakes:

  • Putting a tier that must grow in place on a Cube. The template is immutable; growth is a migration.
  • Picking a CPU family that is not offered in the VDC's location, or one that requires UEFI for an image that boots with BIOS.
  • Counting only disks against the 24 PCI connectors. NICs use them too.
  • Treating a CPU-family change as a live operation. It needs a restart and a window.
  • Creating a Windows server with more than 8 GB RAM from an IONOS CLOUD Windows image.

Summary

Choose the compute class per tier on four axes: exclusive or shared cores, CPU-family control, Block Storage attachment, and the operational model including the SLA. Dedicated Core Servers give exclusive cores and a family you control; vCPU Servers trade both for cost. Cubes attach Block Storage freely but lock their template, and Confidential VM locks its shape and loses live migration in exchange for hardware-enforced isolation. Because no class can be toggled, a wrong choice costs a migration.

Key Points:

  • Compute Engine types cannot be toggled; changing class means moving storage to a new instance.
  • Only Dedicated Core Servers let you choose the CPU family; changing it requires a restart.
  • NICs and storage share 24 PCI connectors on Dedicated Core and vCPU Servers.
  • A Cube's NVMe volume and template are fixed, but it can attach up to 23 add-on Block Storage devices.
  • Confidential VM runs in de/fra/2 with a fixed shape, no live migration, no cloud-init and no startup-volume snapshots.

Important Terminology:

  • Core Technology Choice: the capability to change the CPU family of a provisioned Dedicated Core Server, with a restart.
  • Configuration template: the fixed vCPU, RAM and NVMe size of a Cube, chosen at creation and immutable afterwards.