Minecraft Server on a Mini PC: RAM, Flags and Backups
The server my friends play on is a fanless Intel N100 mini PC that cost $155, sits on a shelf behind the router, and has been up since February apart from one power cut and the week I broke it myself. Eight of us use it.

The server my friends play on is a fanless Intel N100 mini PC that cost $155, sits on a shelf behind the router, and has been up since February apart from one power cut and the week I broke it myself. Eight of us use it. Nobody pays a monthly fee, nobody waits for a queue, and the world folder is now 2.1 GB because one person decided to map every ravine within four thousand blocks.
Most of the advice you find for this assumes you are renting. Rented hosts sell you RAM because RAM is the number they can print on a pricing page, which is how people end up convinced a server for eight friends needs 16 GB. It does not. What it needs is single thread speed, a world that is not asking the main tick loop to do impossible work, and a backup you have restored at least once. Those three things, in that order.
RAM is the wrong number to shop for
A Minecraft server runs its world simulation in one main thread. Mobs, redstone, block updates, and entity movement all happen in sequence inside a tick that the server tries to finish in 50 milliseconds, twenty times a second. Adding cores does not make that tick faster. Adding memory does not either, past the point where the heap stops filling up.
The number to watch is MSPT, milliseconds per tick. Run the tick query command on a vanilla server, or install the Spark profiler on Paper, and you get an average. Under 30 is comfortable. Between 40 and 50 the world still runs at full speed but has no headroom, so one player lighting a large area will push it over. Above 50 the server misses ticks and everything feels like it is wading.
On our box, eight players spread across a world with two active farms sits at about 11 MSPT. The worst reading I have recorded was 68, caused by a flying machine left running in a loaded chunk overnight. No amount of extra RAM would have touched it.
What to allocate, by how many people actually log in
These are heap sizes I have run and measured, on Paper, with the world pregenerated and view distance at 7. Modded packs are a different problem entirely and roughly triple every row.
| Concurrent players | Heap to allocate | Host RAM needed | Notes |
|---|---|---|---|
| 2 to 4 | 2 GB | 4 GB | Runs happily on a Raspberry Pi 5 |
| 5 to 10 | 4 GB | 8 GB | Our setup, never seen the heap exceed 2.6 GB |
| 10 to 20 | 6 GB | 16 GB | Single thread speed becomes the limit, not memory |
| 20 plus | 8 GB | 16 GB | Consider splitting worlds across two processes |
Set the minimum and maximum heap to the same value. On our server that means -Xms4G and -Xmx4G. A heap that grows during play gives the garbage collector a moving target, and the pauses show up as a stutter everyone notices at once. Pair that with G1 collector tuning rather than the defaults: a larger young generation and a 200 millisecond pause target, which you can copy from the flag set the Paper community has maintained for years. The practical effect on our box was cutting the longest pause from 340 milliseconds to around 90.
The advice to just forward port 25565 deserves pushback
Almost every guide ends with the same instruction: open 25565 on the router, hand out your IP address, done. It works. It is also the step that makes your home network a public service, and the consequences are not theoretical.
Within two days of opening that port I had scanner traffic in the logs from four continents. Minecraft server scanning is industrialised. Bots crawl the whole IPv4 space, record version strings and player counts, and publish them. What follows is griefers who try joining with a known exploit for the version you advertised, and occasionally a volumetric flood aimed at your home connection that takes the rest of the household off the internet as collateral.
The better default for a group of friends is a mesh VPN. Put the server and every player on a WireGuard based overlay network, and they connect to a private address that no scanner can reach. Setup takes about fifteen minutes per device and costs nothing at this scale. The tradeoff is honest: every new player needs an account on the mesh and a client installed, which is a real nuisance if your group changes weekly. For a stable group of eight it is the right call, and it means the only thing exposed is nothing at all. If you do decide to forward the port anyway, work through what to check before exposing a home server online first, and turn the whitelist on before the server ever accepts its first connection.
Backups that survive the way servers actually die
Copying the world folder while the server is running produces a file set that may or may not restore. The region files are being written to, and a chunk caught mid write comes back corrupt. I learned this the expensive way in March when a restore gave us a world with a hole where the main base used to be.
The sequence that works is three commands and a copy. Issue save-off to stop the server writing, then save-all flush to force everything to disk, take your snapshot, then save-on. On our world that whole window is under four seconds and players do not notice. Run it from a shell script on a cron job at 05:00, send the snapshot to a deduplicating backup tool, and keep one copy off the machine entirely.
Our retention is seven daily snapshots and four weekly. Because the tool only stores changed blocks, a year of that history occupies 9 GB for a world folder that is 2.1 GB. The thing worth doing that almost nobody does is the restore rehearsal: once a quarter, restore the latest snapshot into a second directory, boot a server on a different port, and walk around. Twice that has caught a backup that had been quietly failing on a permissions error for weeks.
The settings that buy you the most headroom
Four changes in server.properties and one external tool account for nearly all the performance we have. View distance down from 10 to 7 cuts the chunks each player keeps loaded by roughly half, and almost nobody can tell by eye. Simulation distance at 5 means mobs and redstone only run close to players. Both of those reduce the work inside that single tick thread.
Lower the global mob cap from 70 to about 40 per player. Entity counts are the most common cause of a slow tick on a friends and family server, usually because someone built a farm and left it loaded. And pregenerate the world before anyone logs in. A chunk generator plugin will build a 4000 block radius in about forty minutes on an N100, after which nobody triggers terrain generation mid flight, which is the other classic source of spikes.
The running cost stays small. That mini PC idles at 7 watts and peaks near 22 with eight people online, which works out to roughly $14 a year here. If you want to check the arithmetic for your own hardware and tariff, the measured figures in home server power cost per year cover the common mini PC and NAS options, and there are more build notes in the self hosting section.
FAQ
Is a Raspberry Pi fast enough?
A Pi 5 handles four or five players on a vanilla world without complaint. Push past that, or add plugins, and the single thread limit arrives quickly. A used x86 mini PC at the same price is the stronger buy for anything beyond a small group.
Paper or the vanilla server jar?
Paper, in almost every case. It keeps vanilla behaviour close enough that nobody notices, fixes a long list of exploits the official jar leaves open, and gives you the profiling tools you need to diagnose a slow tick. The one reason to stay vanilla is if you are testing behaviour that must match the official build exactly.
How much upload bandwidth does it use?
Less than people expect. Each connected player consumes roughly 100 to 150 kbit per second upstream during normal play, with brief spikes when they move into new terrain. Eight players sit comfortably inside a 10 Mbit upload link.
What breaks most often?
Updates. A version bump that changes world format, or a plugin that has not caught up with the new API, accounts for every outage we have had that was not a power cut. Pin your versions and read the release notes before you pull.
More in Self Hosting
Self Hosting
Dynamic DNS Setup to Reach Your Home Server by Name
The server at home is fine. The router is fine. What broke is the sentence you memorised two months ago, the one with four...
Self Hosting
Updating Self Hosted Containers Without Breaking Them
Nothing on the server is broken, which is exactly why the Nextcloud container on mine spent fourteen months running an image I...
Self Hosting
Authelia Setup to Add Two Factor Login to Home Server Apps
You finally got Jellyfin, Paperless and a dashboard reachable from outside the house, and each one has its own login page, its...