Cloud Server Sizing Guide: How Much CPU, RAM and SSD?

Cloud Server Sizing Guide: How Much CPU, RAM and SSD?

Cloud Server Sizing Guide: How Much CPU, RAM and SSD?

Cloud server sizing starts with the workload, not the hardware. This guide gives you starting ranges of vCPU, RAM and disk for common workloads, maps them onto the Small, Medium and Large presets of the DCXV cloud server page, and shows how to check the result with a few shell commands after deployment. Every range below is an estimate and a starting point, not a measurement. Your code, your traffic and your data decide the real numbers.

Start with the workload

Write down four facts before you open any configurator: what runs (web server, database, Windows application), how many people or requests hit it in the busiest hour, how much data you hold today, and how fast that data grows.

Then find the resource that hurts first. A PHP shop usually runs out of RAM before CPU. A build server runs out of CPU. A database runs out of RAM, then out of disk speed. Size for that bottleneck and let the other resources follow.

Starting ranges by workload

The presets: Small is 2 vCPU, 4 GB RAM, 40 GB NVMe from EUR 15.00/mo. Medium is 4 vCPU, 8 GB RAM, 80 GB NVMe from EUR 18.43/mo. Large is 8 vCPU, 16 GB RAM, 160 GB NVMe from EUR 36.88/mo. Prices checked 2026-10, each price includes one weekly backup. See what drives the price before you compare.

Workload Starting range (estimate) Closest preset
Dev/test, VPN, monitoring, utility 1-2 vCPU, 1-4 GB RAM, 20-40 GB disk Small
Small site or blog, low traffic 2 vCPU, 2-4 GB RAM, 20-40 GB disk Small
Online shop, API or SaaS MVP 2-4 vCPU, 4-8 GB RAM, 40-80 GB disk Medium
App with its own database 4-8 vCPU, 8-16 GB RAM, 80-160 GB disk Medium or Large
Windows/RDP business app Vendor formula first. M.E.Doc terminal: CPU = 4 + users, RAM = users x 2 GB + 4 GB Large up to 4 users, custom above

Assumptions: Linux, one main application, normal traffic with occasional peaks, no GPU work, no heavy batch jobs. An idle dev box needs less. An unoptimised plugin stack needs more. For the Windows row the vendor formula outranks any table: it is a requirement, not an estimate, and it changes with the deployment type, so read the vendor's current page. If your numbers fall between presets or above Large, build a custom configuration in the calculator on the cloud server page.

A worked RAM budget for a small shop

Here is an illustrative budget for one online shop on one server. The figures are assumptions, not measurements.

  • Operating system and services - 0.5 GB.
  • PHP-FPM - 10 workers x 100 MB = 1 GB.
  • Database buffer pool - 2 GB.
  • Redis cache - 0.5 GB.
  • File cache and spikes - 1 GB.

The total is about 5 GB. A 4 GB server would swap under load. An 8 GB server keeps roughly 3 GB, about a third, free. Replace each line with your own numbers once the shop runs: the commands below show them.

Headroom, operating system and storage

  • CPU - aim for sustained load under about 70% at your busiest hour, so a traffic spike does not queue requests.
  • RAM - keep 20-25% available. Constant swapping means the server is too small.
  • Disk - stay under 80% full. Databases, logs and updates grow faster than plans.
  • Operating system - a minimal Linux needs a few hundred MB. Windows Server with a desktop needs a few GB before your application starts (estimate).
  • Disk speed - capacity is not speed. A busy database cares about latency, so measure it.

Backup capacity is not production storage

The price of each preset includes one weekly backup. That backup is a safety copy of the server disk, not extra room for your files. Do not count it when you size the disk, and do not plan to keep archives, media or dumps "in the backup". Production data lives on the server disk, so size that disk for data, growth and headroom. Before a risky change, add a snapshot: see cloud VM snapshots.

Measure after deployment

Measure during a normal busy hour and again during a peak. On Debian or Ubuntu, iostat comes from the sysstat package. To sample RAM for a whole week and get a right-sized figure, sample MemAvailable every few minutes, as in measuring the RAM you really use during the 2026 DRAM squeeze.

top -b -n 1 | head -n 15
free -h
vmstat 5 5
df -h
sudo apt install sysstat
iostat -xz 5 3
ps -eo pid,comm,%mem,%cpu --sort=-%mem \
  | head -n 8
sudo du -xh --max-depth=1 /var \
  | sort -h \
  | tail -n 8

How to read the output:

  • top - compare the load average with your vCPU count. A load above the vCPU count for long periods means CPU pressure. The st value is CPU time taken by the host. A busy CPU can still be waiting on RAM: see memory latency on a cloud server.
  • free - read the available column, not free.
  • vmstat - non-zero si and so mean swapping.
  • iostat - high %util and growing await mean the disk is the bottleneck.

For a whole application rather than one workload, follow the worked example in How to Size a Cloud Server for SaaS: a memory budget per container, a k6 load test and a printable worksheet.

Choose and verify

Take the closest preset from the table and run it under real load. A short test is the honest way to check an estimate: see the trial checklist. Write to sales@dcxv.com to ask about a short trial; it is at DCXV's discretion and not guaranteed. If your measurements do not match the table, trust the measurements. Not sure about your numbers? Send your requirements to sales@dcxv.com and an engineer will review them, or open the cloud server page and pick a preset.

SSD Prices May Fall in 2027: What Server Buyers Should Do
dedicatedcloudpricinginfrastructure

SSD Prices May Fall in 2027: What Server Buyers Should Do

TrendForce sees NAND contract prices falling 0-5% in 1Q27 after a year of steep rises. What it means for server storage, with our own per-TB drive price data.

Starlink Mini 2 Behind CGNAT: Reaching a Site That Moves
networkingcloudinfrastructure

Starlink Mini 2 Behind CGNAT: Reaching a Site That Moves

Starlink Mini 2 is out, but standard plans sit behind CGNAT. We counted 990,208 IPv4 addresses for 12M+ customers and compare four ways to reach a remote site.

Data Center Protests Reach Europe: What Buyers Should Do
clouddedicatedinfrastructure

Data Center Protests Reach Europe: What Buyers Should Do

Hundreds entered Microsoft's Amsterdam data center site on 6 November. The Dutch power numbers behind it, measured latency to EU hubs and what buyers should do.

US Voters Rejected Data Centers 45 to 1: What Buyers Learn
clouddedicatedinfrastructure

US Voters Rejected Data Centers 45 to 1: What Buyers Learn

On 3 November 47 US ballot measures on data centers went to voters and the restrictive side won 45. We recount the raw data: what it means for server buyers.

AWS BYOIP in 2026: $43.80 per IPv4 vs Your Own /24
ipv4pricingcloud

AWS BYOIP in 2026: $43.80 per IPv4 vs Your Own /24

AWS bills $0.005 an hour per public IPv4, $43.80 a year. BYOIP addresses carry no charge: when a bought or leased /24 pays back, with a ROA readiness check.