Methodology
How we plan to test game server hosts
GameServerBench plans to run each host on the same game build, the same save and the same load, on a comparable plan, with dated results and raw logs anyone can inspect. None of that has started: there is no test rig, no test server and no result. This page is the protocol the site intends to follow, written down before any run so it cannot be bent to suit a result.
The protocol
Four documents, in order.
- 1Test rigWhat the measurement runs on, and how it is recorded.
- 2Load profilesWhat the server is asked to do during a run.
- 3Stable playerThe threshold a server has to hold to count a player.
- 4IndependenceWho pays for what, and what host terms allow.
A number appears here only with all six of these attached.
- Method version
- Run date
- Host and plan
- Game and version
- Load profile
- Raw log
What has to happen first
Five things stand between this plan and a published number:
- Budget. Spending on test servers has not been approved.
- A working harness. The logging and load tooling has not been built. It would be proven on owned hardware first, to show that two identical runs agree.
- Permission. Each host's terms must be checked, and written confirmation sought where they are unclear. See independence and host terms.
- Fixed definitions. The thresholds in the stable player definition must be locked before the first run.
- Purchase rules. Any test server would be bought at the public price, without an affiliate link.
What a fair comparison would hold constant
A difference between hosts means little unless everything else stayed the same. Each run would record, and hold fixed where possible:
| Held constant | Why it matters |
|---|---|
| Game build, modpack version, loader and Java version | A patch can change performance more than a host does |
| Save or world copy, restored before each run | World age and size drive memory and tick cost |
| Load script or recorded session | Idle bots do not do what players do |
| Plan tier and allocated RAM | Comparing unequal plans would measure price, not the host |
| Region and client location | Distance changes latency and has nothing to do with the host's hardware |
| Date and time of day | Shared hardware can vary with what neighbours are doing |
The test rig plan covers tools and locations, and the load profiles cover what simulated or recorded play would look like for each game.
What would be measured
The core reading is how long the server takes to do its work, using the game's own reporting wherever possible. On Minecraft that is MSPT and TPS, which spark's documentation defines around a 50 ms budget per tick, and spark is bundled with Paper from 1.21. Palworld exposes serverfps and serverframetime through its REST API metrics endpoint. Averages hide lag spikes, so results would report median, 95th percentile and worst case. See MSPT vs TPS for what those mean.
Memory use over time, boot time, join success and round-trip time from named locations would be recorded beside tick time. Three separated runs would stand behind any headline number, because one run cannot separate a host from a bad hour.
How long a run would be
Most comparisons need hours, and a few need weeks. Long-run claims about memory growth or month-over-month degradation need a 30-day run on each host. In the current plan that applies to two pages: Palworld and All the Mods 10. Neither makes such a claim today.
How results would be published
Each result would carry its date, build, plan, region and run count, and link to its raw logs on the data page. Retired plans would stay as dated archive rows, and a host could use its right of reply after publication.
Why publish a method with no data
Describing a test is not the same as showing one, and buyers have little way to tell the difference. The published evidence GameServerBench has found is thin:
- Godlike publishes per-host Minecraft TPS tables against several competitors, including its method. Godlike sells hosting and is a party in each comparison, and it describes the work as a single controlled run and not a long-term study.
- MamboServer's Palworld roundup (updated 7 January 2026) describes a test process and reports CPU and disk figures for one host, Godlike. It shows no per-host tick tables, and it discloses referral income.
GameServerBench found no independent, non-selling publisher with per-host game tick data. You can collect your own data on a server you rent: how to benchmark your own game server shows how, starting with a check of your host's terms.