GUIDE · PRIVATE CLOUD

Capacity planning for virtualisation: how to size CPU, RAM and storage

In this guide you will learn how to estimate the resources a virtualised environment needs, starting from real workloads rather than impressions. After reading it you will know how to collect the right data, apply a reasoned margin, decide what to verify before buying and recognise the errors that lead to oversized or undersized servers.

Updated october 2026
8 min read
Altovar Technical Editorial Team

1. Why wrong sizing is costly

An undersized server slows down every virtual machine it hosts and turns a normal peak into an emergency. An oversized server makes the opposite mistake: it ties up capital, increases software licences that depend on processors and consumes energy without any real benefit. Capacity planning serves to avoid both scenarios by linking resources to what the systems actually do over time.

The first mistake to avoid is confusing installed capacity with actual load. A server with a great deal of memory can be saturated if that memory is already occupied by active systems, while an environment whose processors are lightly used is a good candidate for consolidating several systems onto fewer machines. For this reason the first question is not how many resources to buy, but how many the systems use today, at what times and with what margin. Without measured data, every choice remains an estimate based on the impression of whoever makes it.

2. What you need before you start

Before calculating any figures you need an inventory. For each system you must know what it does, which users rely on it, at what times and with what tolerance for downtime. You also need performance measurements collected over a period long enough to include normal and peak times, such as period-end accounting, sales campaigns or overnight processing. A single day of measurements represents only that day.

  • List of systems with owner, business function and hours of use.
  • History of processor, memory, disk and network usage collected over representative periods.
  • Calendar of known peaks, with the dates and the activities that generate them.
  • Continuity requirements for each service, that is, how long it may stay down.
  • Software licences that depend on the number of processors or machines.

3. The steps of sizing

Sizing is a repeatable process, not a single formula. Start with the most critical systems and those whose load changes over time, because they create the most surprises. Each step produces a number to record, together with the assumption that supports it. If an assumption is not documented, nobody will be able to verify it when the load changes, and the calculation will have to be redone from scratch.

  1. Group the systems by load profile: those that use a lot of CPU, those that depend on RAM, those that write heavily to disk.
  2. For each system record the typical value and the peak value, not just the average, because the average hides critical moments.
  3. Add together the loads that can coexist on the same server, considering peaks that coincide in the same time interval.
  4. Apply a growth margin agreed with the owners of the services, based on concrete forecasts rather than habit.
  5. Add the capacity needed for redundancy, that is, the ability to host the systems of a failed server on the others.
  6. Compare the result with the available configurations and record the assumptions used, together with the periods observed.

4. Turning the data into CPU, RAM and storage

Each resource follows different logic. CPU is exhausted at peak times and must be read carefully against the hours of the day. RAM can be allocated to virtual machines in amounts greater than the physical memory, but not without limit, because the system must remain responsive when many systems are working together. Storage must be assessed for both capacity and performance, because a disk with plenty of free space can be slow precisely when it is most needed.

ResourceWhat to measureTypical error
CPU (processor)Average and peak utilisation per system, with application response timesRelying on average utilisation and ignoring overnight or end-of-period processing
RAM (working memory)Memory actually active, not the memory assigned to virtual machinesAssigning memory to every machine without considering what they really use
Storage (disks)Space occupied, monthly growth and read and write performanceCalculating only capacity and forgetting the speed databases require

5. Growth margins and redundancy

The growth margin protects against the fact that loads almost always increase. There is no percentage valid for everyone: it depends on how fast the business is growing and on the projects planned for the coming years. Redundancy must also be considered separately. If one server must be able to host the systems of another that fails, the usable capacity of the group is lower than the installed capacity, and this must be stated when the choice is made, not after the first failure.

6. Common mistakes to avoid

The most frequent errors are not technical but methodological. Sizing is done over too short a period, coinciding peaks are ignored, and manual resource allocation is allowed to grow without control. Use this list as a check at every review, because each point is easy to forget when the environment changes. The criterion is simple: if an item has not been verified, the conclusion is not yet valid.

Each item in the list signals a question to put to the team and to the provider before approving a configuration. If the answer is unclear, the configuration should not be approved until the doubt is resolved. Waiting a few days is better than correcting an error with the system already in production and users connected. These are the simplest checks to carry out and the most overlooked.

  • Using only one month of data, or only one day, without considering annual cycles.
  • Assigning each virtual machine its maximum expected size, creating locked and unused capacity.
  • Forgetting backup and replication processes, which consume resources precisely at night.
  • Overlooking processor-based licence costs when choosing a configuration.
  • Failing to review sizing after a change of software or in the number of users.

7. How to check that the sizing holds

Sizing should be verified in two ways. The first is comparison with real data after go-live: if the systems stay consistently below the agreed thresholds even at peak times, the assumption holds. The second is a controlled test, for example switching off one server in an agreed window and observing whether the other systems can carry the moved load. Record the results in writing, because they will become the starting point for the next review.

8. How to start: the first step this week

The first concrete step is to prepare the inventory of systems with each owner. A full project is not needed: a table with service name, function, hours, downtime tolerance and point of contact is enough. If the company has many systems, start with those that stop daily work. It is a task of a few hours, and it makes the discussion with the provider far more useful, moving the conversation from impressions to data.

Immediately afterwards, start collecting measurements on the most critical systems, including at least one known peak period. When you have both, ask the provider for a sizing proposal with its assumptions written down, and compare it with what you have observed. The differences between the two readings are the most valuable part of the work. This step takes at most a few days and involves no purchases, so it is the cheapest way to avoid wrong decisions.

Frequently Asked Questions

How long should I collect data before sizing?

There is no fixed duration. The aim is to cover a period in which typical loads and known peaks appear, such as a period-end close or a busy season. Too short a period leads to optimistic estimates. If activity is very stable you can start with a shorter window and extend it over time, but always document which period you observed.

Is it wise to oversize for safety?

A margin is necessary, but a blanket excess is costly and often pointless. Better a declared, justified margin for each service, with periodic checks of its use. Unused resources protect nothing: they become licence costs, energy consumption and occupied space. If a service is critical, the answer is usually redundancy rather than excess power on a single server.

Does sizing need to be repeated over time?

Yes, and it is one of the most neglected tasks. Sizing goes stale when applications change, the number of users grows, the amount of data to retain increases or backup processes change. A periodic review, linked to the times when new projects are planned, avoids surprises. Set a regular appointment and review after every significant change.

Who should do the sizing: me or the provider?

Responsibility remains yours, because you know the workloads and the business priorities. The provider brings the method, measurement tools and experience with configurations. The most robust solution is joint work: you supply the inventory and downtime tolerances, and the provider proposes the configuration with written assumptions and checks. Be wary of anyone who quotes a figure without having seen the data.

Does public cloud remove the need to size?

It reduces some problems, because capacity can be added more flexibly, but it does not eliminate them. Usage profiles, variable costs and contractual limits still have to be decided. Even in a private or hybrid infrastructure the measurement work is the same.

About this content

General informational guide. Regulations and technical requirements change over time: always check the official sources and assess your specific case with a qualified technical or legal advisor.

GET IN TOUCH

Talk to our team.

We'll show how Altovar integrates into your infrastructure. No commitment, no sales pitch. Just a technical conversation.

BOOK A DEMO →