How to Benchmark Your Linux Server (Disk, CPU, Memory and Network — and What the Numbers Mean)
"The server feels slow" and "would a mini PC be faster?" are both answered with numbers, and the tools are free: fio for disks, sysbench for CPU and memory, iperf3 for the network (127). Run them once on a fresh machine and you have a baseline; run them again when things feel off and you know whether it's the hardware or the software. The typical results below are for comparison.
Quick reference
| What | Command | Good home-server number | |
|---|---|---|---|
| Disk, sequential read | fio --name=r --rw=read --bs=1M --size=2G --direct=1 | SATA SSD 500 MB/s · NVMe 2000+ · HDD 150–250 | |
| Disk, random 4K | fio --name=rr --rw=randread --bs=4k --size=1G --direct=1 --iodepth=32 | SSD 50k+ IOPS · HDD 100–300 · SD card 1–3k | |
| CPU | sysbench cpu --threads=$(nproc) run | events/s: Pi 4 ≈ 2,500 · N100 ≈ 7,000 · i5-8400 ≈ 10,000 | |
| Memory | sysbench memory run | GB/s: Pi 4 ≈ 3 · N100 ≈ 10 · desktop DDR4 ≈ 15+ | |
| Network | iperf3 -c server | gigabit ≈ 940 Mbps · 2.5G ≈ 2,350 | |
| Everything, one score | `curl -sL yabs.sh | bash` | compare to other people's posted results |
1. Disk with fio
sudo apt install fio -y
cd /mnt/data # test the drive you care about
Sequential (big files: media, backups):
fio --name=seqread --rw=read --bs=1M --size=2G --numjobs=1 --direct=1 --group_reporting
fio --name=seqwrite --rw=write --bs=1M --size=2G --numjobs=1 --direct=1 --group_reporting
Random 4K (databases, Immich, Docker layers — what "feels slow" usually is):
fio --name=randread --rw=randread --bs=4k --size=1G --numjobs=4 --iodepth=32 --direct=1 --group_reporting
fio --name=randwrite --rw=randwrite --bs=4k --size=1G --numjobs=4 --iodepth=32 --direct=1 --group_reporting
Read bw= (throughput) and IOPS=. Delete the test files after (rm seqread* randread*). On ZFS, --direct=1 may be refused; drop it and note the result includes cache.
What it tells you: a system disk under 10k random IOPS (an SD card, a USB stick, a 2.5" HDD) is why containers start slowly and Nextcloud lags; moving Docker and databases to an SSD is the single biggest upgrade (tutorial 38 for the Pi).
2. CPU and memory with sysbench
sudo apt install sysbench -y
sysbench cpu --cpu-max-prime=20000 --threads=$(nproc) run
sysbench cpu --cpu-max-prime=20000 --threads=1 run
sysbench memory --memory-block-size=1M --memory-total-size=10G run
Read events per second (CPU) and MiB/sec (memory). The single-thread number predicts how snappy web apps feel; the all-threads number predicts transcoding and photo indexing.
Rough reference, events/s all threads: Raspberry Pi 4 ≈ 2,500; Pi 5 ≈ 6,000; Intel N100 mini PC ≈ 7,000; an eight-year-old i5 desktop ≈ 10,000; a modern i5/Ryzen 5 ≈ 25,000+. Memory: Pi 4 ≈ 3 GB/s; N100 ≈ 10; DDR4 desktop ≈ 15–20.
3. Network with iperf3
Tutorial 127 covers it: iperf3 -s on the server, iperf3 -c 192.168.1.50 -t 20 from a wired client. Gigabit should show ≈ 940 Mbps; if you see 94, something's negotiating at 100 Mb/s (117). Test Wi-Fi from the rooms that matter, and -R for the other direction.
4. The whole thing at once
YABS (Yet Another Benchmark Script) runs fio, Geekbench and an internet speed test and prints a shareable table:
curl -sL yabs.sh | bash -s -- -i # -i skips the internet test
Save the output in BookStack (88) as the machine's baseline; the Geekbench score compares directly with any CPU on browser.geekbench.com.
5. Real-world checks
Synthetic numbers are a start; these are what you actually notice:
- time docker compose up -d on a stopped stack — disk IOPS and CPU together.
- time tar -czf /dev/null /home/john/docker — CPU compression speed.
- Jellyfin: how many 1080p transcodes before stutter (Dashboard shows them). CPU, or GPU if VAAPI (42) is on.
- Immich: photos indexed per minute on first import — CPU, and a reason for a GPU.
- Nextcloud: the admin page's "slow" warning means the database disk.
6. When the numbers say "upgrade"
- Random 4K IOPS under 5k on the system disk → SSD, before anything else.
- Single-thread CPU under 1,000 events/s and you run Jellyfin/Immich/Nextcloud → a mini PC (119 has the power maths).
- Memory bandwidth fine but free -h always short → RAM, cheaper than anything.
- Network at 940 Mbps but Samba copies at 60 MB/s → the disk, not the network.
- Everything fast in benchmarks, server still slow → software: tutorial 16 and 68, not hardware.
Troubleshooting
fio numbers wildly high. Cache: missing --direct=1, or a size smaller than RAM on a filesystem that caches. Use a bigger --size or direct I/O.
sysbench not found on a Pi. Older repos; pipx install sysbench or build from source, or use openssl speed as a rough CPU check.
Results vary run to run. Other things were running; stop Docker stacks for a clean baseline, and run three times.
Benchmark heated the box and it throttled. That's a finding (117, 119): check sensors during the run.
Wrapping up
fio for the disk, sysbench for CPU and memory, iperf3 for the wire, YABS for a shareable table. Numbers in the wiki when the machine is new, numbers again when it feels wrong, and the table above to say whether the hardware is the problem — or, usually, that an SSD is.
