Skip to content
5m Read

How to benchmark a VPS

How to benchmark a VPS
Get an AI summary

A benchmark tells you how good your VPS actually is; from CPU to disk and network. This guide walks through how to run standard testing with one command, how to read what it prints, and how to get numbers you can fairly compare across providers.

The short version

Log in to the server as any user and run:

Bash
1curl -sL yabs.sh | bash

This runs YABS (Yet Another Bench Script), the script most of the VPS community uses to compare servers. It needs no root access and installs nothing permanently. It tests three things:

Test

Tool

What it measures

Disk

fio

Random read and write speed at 4k, 64k, 512k and 1m block sizes, 50% reads and 50% writes

Network

iperf3

Download and upload speed to public test servers in several regions

CPU and Memory

Geekbench 6

Single-core and multi-core scores you can compare with any other machine

A full run takes a few minutes. The network test pushes the port as hard as it can for about 20 seconds per location, so don't run it on a server that's serving production traffic.

Read the script before you run it

Piping a script from the internet straight into runs whatever it contains. YABS is widely used and open source, but you should still read it first on any server that matters:

Bash
1curl -sL yabs.sh -o yabs.sh
2less yabs.sh
3bash yabs.sh

YABS downloads prebuilt fio and iperf3 binaries, and a Geekbench tarball, at run time. If fio or iperf3 is already installed, YABS uses the local copy instead.

01-download-yabs-1.png

What to measure and why

Each part of the result answers a different question about the server.

  • Single-core CPU decides how fast one request, build step or query runs. Most web apps and databases care about this number more than any other.
  • Multi-core CPU decides how much parallel work the server can do at once, such as builds, video encoding or many concurrent users.
  • 4k random disk shows how the disk handles the small, scattered reads and writes a database makes. It's the best single sign of whether the storage is fast or contended.
  • 1m disk shows sequential throughput: backups, large file copies and media.
  • Network shows how fast the server can move data to the rest of the internet, which matters for downloads, backups and serving large files.
  • Consistency is the number most benchmarks skip. A server that scores well once and poorly an hour later has neighbours competing for the same hardware.

Choose what to test with flags

Pass flags after to skip tests or change what runs:

Bash
1curl -sL yabs.sh | bash -s -- -i

Flag

What it does

-i

Skips the network test. Use this on metered or low-bandwidth servers.

-r

Tests against fewer network locations, to use less bandwidth.

-f

Skips the disk test.

-g

Skips Geekbench.

-7

Runs Geekbench 7 instead of Geekbench 6.

-j

Prints the results as JSON as well.

-w results.json

Writes the JSON results to a file.

You can combine flags. skips both disk and Geekbench, so only the network test runs.

Geekbench 6 is the default because most published results use it. Only compare a Geekbench 6 score with other Geekbench 6 scores; versions score differently.

Read the output

02-yabs-run1.png

Here's what each section tells you.

Basic System Information lists the CPU model, core count, RAM and disk the server reports. Check the CPU model against what the provider advertises. shows the virtualization, such as .

fio Disk Speed Tests shows throughput and IOPS at each block size. Look at the 4k column first. Compare it with results from other servers of the same price and disk type rather than an absolute number: NVMe on a quiet host should beat SATA SSD, and both should be far ahead of network storage.

iperf3 Network Speed Tests shows send and receive speed per location. Results drop with distance, so a US server is slower to Europe than to New York. A busy test server is skipped and shows , which isn't a fault on your side.

Geekbench 6 shows single-core and multi-core scores, plus a link to the full result on the Geekbench website. Open the link and compare your score with others from the same CPU model. A score well below others on the same CPU suggests the host is oversubscribed or the cores are shared.

Go deeper

YABS gives you a broad picture. When one part of it matters most for your workload, test that part directly.

05-geekbench-with-swap.png

Database-style disk load

This fio job mimics a database: random 4k reads and writes, 70% reads, with several requests in flight. Run it in a directory on the disk you want to test:

Bash
1fio --name=dbtest --rw=randrw --rwmixread=70 --bs=4k --size=2G \
2 --iodepth=32 --numjobs=4 --direct=1 --runtime=60 --time_based \
3 --group_reporting --ioengine=libaio
4rm -f dbtest.*

bypasses the page cache, so you measure the disk rather than RAM. Read the `IOPS` line for reads and writes, and the `clat` percentiles for latency. A high 99th percentile latency hurts a database more than a low average IOPS.

CPU steal

On shared CPU plans, the host divides physical cores between several servers. Steal is the share of time your server wanted the CPU and had to wait. Watch it while the server is busy:

Bash
1vmstat 1 30

The column is steal, as a percentage. Values that stay near zero are normal. Steal that is regularly above a few percent means other tenants are competing for your cores, and your app will feel slower at those times. (from the package) shows the same figure as .

Dedicated CPU plans don't share cores, so steal stays at zero. If your workload is latency-sensitive, compare a run on a shared VPSwith one on a dedicated-core VDS.

Network to your own users

The public iperf3 servers YABS uses tell you about the provider's upstream, not the path to your users. To test that path, run an iperf3 server on a machine near your users and test against it:

On a machine near your users:

Bash
1iperf3 -s

On the VPS:

Bash
1iperf3 -c <that machine's IP> -P 4 -t 30
2iperf3 -c <that machine's IP> -P 4 -t 30 -R

reverses the direction, so the second command measures the VPS's download speed. For latency alone, shows the delay and loss at each hop.

Get results you can trust

One run is a snapshot. To compare providers fairly:

  • Test an idle server. Stop your own services first, or test a fresh server before you deploy anything.
  • Run it at least three times, at different hours. Neighbours on shared hardware are busiest at different times of day. Report the median and note how far the runs spread.
  • Compare like with like. Use the same plan size, region type, virtualisation and Geekbench version.
  • Note the region. Network results depend heavily on where the server and test endpoints are.
  • Save the JSON. keeps the raw numbers, so you can compare later runs without retyping anything.

Compare your results

Once you have numbers, compare them against other servers:

- YABSdb and VPSBenchmarks collect YABS results from many providers. YABS can submit to either with the flag.

- The LowEndTalk community has long-running threads where people post YABS results by provider.

- Our benchmark race runs the same tests on a similar-priced plan at several providers, including ours, on a schedule.

Frequently Asked Questions


News & Insights