Key results
-
80 bots
Paper with Aikar flags
Averaged over the measurement window, Paper stayed, up to 80 bots in all three repetitions, below the limit of 50 ms p95 MSPT fixed in advance, and in one of them up to 85. With the stricter limit of at least 19.5 TPS on average, it was 80 in all three repetitions.
Evidence: laeufe.csv, Field kapazitaet · laeufe.csv, Field kapazitaet_tps · Source [1]
-
75 bots
Purpur, not separable from Paper at capacity
Purpur reached 75 to 80 bots. The range overlaps with that of Paper (80 to 85), so this measurement shows no difference between the two.
Evidence: laeufe.csv, Field kapazitaet
-
65 bots
Fabric with Lithium
Fabric with Lithium reached 65 to 70 bots. That is above Vanilla and below Paper and Purpur; the ranges do not overlap in either direction.
Evidence: laeufe.csv, Field kapazitaet
-
55 bots
Vanilla
Averaged over the measurement window, Vanilla stayed, in all three repetitions and up to 55 bots, below 50 ms p95 MSPT. With an average of at least 19.5 TPS over the measurement window, it was only 50.
Evidence: laeufe.csv, Field kapazitaet · laeufe.csv, Field kapazitaet_tps
-
16.31 TPS
The MSPT limit alone is not enough
At 55 bots, Vanilla was below 50 ms p95 MSPT on average over the measurement window, but ran at only 16.31 to 17.16 instead of 20 ticks per second. This is why this study reports the capacity at at least 19.5 TPS alongside the capacity fixed in advance.
Evidence: daten.csv, Field tps · Source [2]
-
80 bots
Default G1 and Aikar flags: no demonstrated capacity difference
Paper with default G1 reached 80 to 85 bots, and with the Aikar flags likewise 80 to 85. There is no demonstrated difference in capacity. Observed in addition: for the p99 value, the Aikar flags were separately lower in 7 of 13 shared levels, default G1 in none. On average, the Aikar runs came earlier in the run order, so part of the gap may stem from drift over the measurement night (see limitations).
Evidence: laeufe.csv, Field kapazitaet · p99-aikar-g1.csv, Field spieler · Source [1]
-
70 bots
Generational ZGC came in lower
Paper with Generational ZGC reached 70 bots in all three repetitions, and thus fewer than with default G1 or the Aikar flags (80 to 85 each). This applies to this load profile, this heap and this Java version.
Evidence: laeufe.csv, Field kapazitaet · Source [3]
-
18 of 18 runs
All runs valid
All 18 runs met the rules fixed in advance. Averaged over a level, CPU steal reached at most 0.061%, and the load generator's cores at most 5.72%.
Evidence: laeufe.csv, Field lauf · daten.csv, Field steal_pct · daten.csv, Field cpu_bots_pct
Summary
Research question: How many concurrent protocol bots can Vanilla, Paper, Purpur and Fabric with Lithium handle on the same hardware, and how strongly do JVM settings affect the result?
We put Vanilla, Paper, Purpur and Fabric with Lithium under load in Minecraft 26.3 on the same virtual machine. Protocol bots joined in stages until the server was overloaded. All 18 runs were valid under the rules fixed in advance.
Capacity is the highest number of bots up to which the p95 tick time stayed below 50 ms on average over the measurement window, in all three repetitions. Paper reached 80, Purpur 75, Fabric with Lithium 65 and Vanilla 55 bots. Paper and Purpur cannot be separated at the capacity defined in advance; for all other pairs of programs, the ranges of the repetitions do not overlap.
For the JVM settings, no demonstrated capacity difference appeared on Paper between default G1 and the Aikar flags. Generational ZGC came in below them at 70 bots.
Tick time alone does not reveal every overload: at 55 bots, Vanilla stayed below 50 ms p95 on average, but reached only 16.31 to 17.16 instead of 20 ticks per second. This is why each capacity is accompanied by the value at an average of at least 19.5 TPS. The figures apply to this VM and this bot profile and cannot be transferred directly to real players.
Results
How to read the results
The capacity fixed in advance is based on the p95 tick time. The second capacity instead requires that the server manages at least 19.5 of 20 ticks per second on average over the measurement window and that all bots stay connected. Both report the smallest value from three repetitions. A difference counts only if the ranges of the repetitions do not overlap.
Capacity of the programs
Highest bot count with below 50 ms p95 MSPT on average over the measurement window in all three repetitions, all with Aikar flags.
- Paper, Aikar-Flags80 Bots
- Purpur, Aikar-Flags75 Bots
- Fabric + Lithium, Aikar-Flags65 Bots
- Vanilla, Aikar-Flags55 Bots
Capacity of the programs at at least 19.5 TPS
Added afterwards: highest bot count with at least 19.5 TPS on average in all three repetitions.
- Paper, Aikar-Flags80 Bots
- Purpur, Aikar-Flags75 Bots
- Fabric + Lithium, Aikar-Flags65 Bots
- Vanilla, Aikar-Flags50 Bots
Capacity of Paper per JVM profile
Highest bot count with below 50 ms p95 MSPT on average over the measurement window in all three repetitions.
- Paper, Aikar-Flags80 Bots
- Paper, Standard-G180 Bots
- Paper, Generational ZGC70 Bots
p95 tick time per level
Each point is one level of a run, shown as the mean of the measurement window. The limit of 50 ms applies to this mean.
- Fabric + Lithium, Aikar-Flags
- Paper, Aikar-Flags
- Paper, Generational ZGC
- Paper, Standard-G1
- Purpur, Aikar-Flags
- Vanilla, Aikar-Flags
Actual tick rate per level
Each point is one level of a run. The target is 20 ticks per second.
- Fabric + Lithium, Aikar-Flags
- Paper, Aikar-Flags
- Paper, Generational ZGC
- Paper, Standard-G1
- Purpur, Aikar-Flags
- Vanilla, Aikar-Flags
All combinations
| Combination | Capacity | Highest run | At 19.5 TPS | Highest run (TPS) | p95 at 50, min | p95 at 50, max | p99 at 50, min | p99 at 50, max |
|---|---|---|---|---|---|---|---|---|
| Paper, Aikar-Flags | 80 | 85 | 80 | 80 | 27.8 ms | 29.8 ms | 31.2 ms | 33.4 ms |
| Paper, Standard-G1 | 80 | 85 | 80 | 85 | 28.3 ms | 29.5 ms | 34.4 ms | 36.1 ms |
| Purpur, Aikar-Flags | 75 | 80 | 75 | 80 | 29.4 ms | 31.0 ms | 33.2 ms | 35.0 ms |
| Paper, Generational ZGC | 70 | 70 | 70 | 70 | 34.1 ms | 35.4 ms | 39.9 ms | 40.3 ms |
| Fabric + Lithium, Aikar-Flags | 65 | 70 | 65 | 65 | 36.1 ms | 37.3 ms | 43.4 ms | 45.0 ms |
| Vanilla, Aikar-Flags | 55 | 55 | 50 | 50 | 42.7 ms | 43.5 ms | 50.2 ms | 55.1 ms |
All 18 runs
| No. | Combination | Rep. | Capacity | TPS there | At 19.5 TPS | Valid |
|---|---|---|---|---|---|---|
| 1 | Paper, Aikar-Flags | 2 | 85 | 19.40 | 80 | Ja |
| 2 | Paper, Aikar-Flags | 1 | 80 | 19.74 | 80 | Ja |
| 3 | Paper, Generational ZGC | 1 | 70 | 20.00 | 70 | Ja |
| 4 | Purpur, Aikar-Flags | 1 | 80 | 19.40 | 75 | Ja |
| 5 | Fabric + Lithium, Aikar-Flags | 2 | 70 | 19.30 | 65 | Ja |
| 6 | Purpur, Aikar-Flags | 2 | 75 | 20.00 | 75 | Ja |
| 7 | Paper, Standard-G1 | 2 | 85 | 19.30 | 80 | Ja |
| 8 | Vanilla, Aikar-Flags | 3 | 55 | 16.81 | 50 | Ja |
| 9 | Vanilla, Aikar-Flags | 1 | 55 | 16.31 | 50 | Ja |
| 10 | Fabric + Lithium, Aikar-Flags | 3 | 65 | 20.00 | 65 | Ja |
| 11 | Paper, Standard-G1 | 3 | 85 | 19.56 | 85 | Ja |
| 12 | Vanilla, Aikar-Flags | 2 | 55 | 17.16 | 50 | Ja |
| 13 | Fabric + Lithium, Aikar-Flags | 1 | 65 | 20.01 | 65 | Ja |
| 14 | Paper, Standard-G1 | 1 | 80 | 20.01 | 80 | Ja |
| 15 | Paper, Generational ZGC | 3 | 70 | 19.82 | 70 | Ja |
| 16 | Purpur, Aikar-Flags | 3 | 80 | 19.57 | 80 | Ja |
| 17 | Paper, Aikar-Flags | 3 | 80 | 19.93 | 80 | Ja |
| 18 | Paper, Generational ZGC | 2 | 70 | 19.97 | 70 | Ja |
Pairwise comparisons
A difference counts as demonstrated if the ranges of the three repetitions do not overlap.
| Metric | A | A min | A max | B | B min | B max | Demonstrated |
|---|---|---|---|---|---|---|---|
| kapazitaet_p95 | Vanilla, Aikar-Flags | 55.0 | 55.0 | Paper, Aikar-Flags | 80.0 | 85.0 | Ja |
| kapazitaet_tps | Vanilla, Aikar-Flags | 50.0 | 50.0 | Paper, Aikar-Flags | 80.0 | 80.0 | Ja |
| p95_bei_50 | Vanilla, Aikar-Flags | 42.7 | 43.5 | Paper, Aikar-Flags | 27.8 | 29.8 | Ja |
| p99_bei_50 | Vanilla, Aikar-Flags | 50.2 | 55.1 | Paper, Aikar-Flags | 31.2 | 33.4 | Ja |
| kapazitaet_p95 | Vanilla, Aikar-Flags | 55.0 | 55.0 | Purpur, Aikar-Flags | 75.0 | 80.0 | Ja |
| kapazitaet_tps | Vanilla, Aikar-Flags | 50.0 | 50.0 | Purpur, Aikar-Flags | 75.0 | 80.0 | Ja |
| p95_bei_50 | Vanilla, Aikar-Flags | 42.7 | 43.5 | Purpur, Aikar-Flags | 29.4 | 31.0 | Ja |
| p99_bei_50 | Vanilla, Aikar-Flags | 50.2 | 55.1 | Purpur, Aikar-Flags | 33.2 | 35.0 | Ja |
| kapazitaet_p95 | Vanilla, Aikar-Flags | 55.0 | 55.0 | Fabric + Lithium, Aikar-Flags | 65.0 | 70.0 | Ja |
| kapazitaet_tps | Vanilla, Aikar-Flags | 50.0 | 50.0 | Fabric + Lithium, Aikar-Flags | 65.0 | 65.0 | Ja |
| p95_bei_50 | Vanilla, Aikar-Flags | 42.7 | 43.5 | Fabric + Lithium, Aikar-Flags | 36.1 | 37.3 | Ja |
| p99_bei_50 | Vanilla, Aikar-Flags | 50.2 | 55.1 | Fabric + Lithium, Aikar-Flags | 43.4 | 45.0 | Ja |
| kapazitaet_p95 | Paper, Aikar-Flags | 80.0 | 85.0 | Purpur, Aikar-Flags | 75.0 | 80.0 | Nein |
| kapazitaet_tps | Paper, Aikar-Flags | 80.0 | 80.0 | Purpur, Aikar-Flags | 75.0 | 80.0 | Nein |
| p95_bei_50 | Paper, Aikar-Flags | 27.8 | 29.8 | Purpur, Aikar-Flags | 29.4 | 31.0 | Nein |
| p99_bei_50 | Paper, Aikar-Flags | 31.2 | 33.4 | Purpur, Aikar-Flags | 33.2 | 35.0 | Nein |
| kapazitaet_p95 | Paper, Aikar-Flags | 80.0 | 85.0 | Fabric + Lithium, Aikar-Flags | 65.0 | 70.0 | Ja |
| kapazitaet_tps | Paper, Aikar-Flags | 80.0 | 80.0 | Fabric + Lithium, Aikar-Flags | 65.0 | 65.0 | Ja |
| p95_bei_50 | Paper, Aikar-Flags | 27.8 | 29.8 | Fabric + Lithium, Aikar-Flags | 36.1 | 37.3 | Ja |
| p99_bei_50 | Paper, Aikar-Flags | 31.2 | 33.4 | Fabric + Lithium, Aikar-Flags | 43.4 | 45.0 | Ja |
| kapazitaet_p95 | Purpur, Aikar-Flags | 75.0 | 80.0 | Fabric + Lithium, Aikar-Flags | 65.0 | 70.0 | Ja |
| kapazitaet_tps | Purpur, Aikar-Flags | 75.0 | 80.0 | Fabric + Lithium, Aikar-Flags | 65.0 | 65.0 | Ja |
| p95_bei_50 | Purpur, Aikar-Flags | 29.4 | 31.0 | Fabric + Lithium, Aikar-Flags | 36.1 | 37.3 | Ja |
| p99_bei_50 | Purpur, Aikar-Flags | 33.2 | 35.0 | Fabric + Lithium, Aikar-Flags | 43.4 | 45.0 | Ja |
| kapazitaet_p95 | Paper, Aikar-Flags | 80.0 | 85.0 | Paper, Standard-G1 | 80.0 | 85.0 | Nein |
| kapazitaet_tps | Paper, Aikar-Flags | 80.0 | 80.0 | Paper, Standard-G1 | 80.0 | 85.0 | Nein |
| p95_bei_50 | Paper, Aikar-Flags | 27.8 | 29.8 | Paper, Standard-G1 | 28.3 | 29.5 | Nein |
| p99_bei_50 | Paper, Aikar-Flags | 31.2 | 33.4 | Paper, Standard-G1 | 34.4 | 36.1 | Ja |
| kapazitaet_p95 | Paper, Aikar-Flags | 80.0 | 85.0 | Paper, Generational ZGC | 70.0 | 70.0 | Ja |
| kapazitaet_tps | Paper, Aikar-Flags | 80.0 | 80.0 | Paper, Generational ZGC | 70.0 | 70.0 | Ja |
| p95_bei_50 | Paper, Aikar-Flags | 27.8 | 29.8 | Paper, Generational ZGC | 34.1 | 35.4 | Ja |
| p99_bei_50 | Paper, Aikar-Flags | 31.2 | 33.4 | Paper, Generational ZGC | 39.9 | 40.3 | Ja |
| kapazitaet_p95 | Paper, Standard-G1 | 80.0 | 85.0 | Paper, Generational ZGC | 70.0 | 70.0 | Ja |
| kapazitaet_tps | Paper, Standard-G1 | 80.0 | 85.0 | Paper, Generational ZGC | 70.0 | 70.0 | Ja |
| p95_bei_50 | Paper, Standard-G1 | 28.3 | 29.5 | Paper, Generational ZGC | 34.1 | 35.4 | Ja |
| p99_bei_50 | Paper, Standard-G1 | 34.4 | 36.1 | Paper, Generational ZGC | 39.9 | 40.3 | Ja |
Method
- Fixing the programsAll programs ran on Minecraft 26.3. Java, server programs and Lithium were verified against the published checksum. No published checksum exists for the Fabric starter, so its own SHA-256 value is recorded. The server.properties were identical for all programs; the remaining settings stayed at each program's defaults. The only other change concerns Paper and Purpur: the per-IP-address connection throttle was switched off, because all bots come from the same loopback address. Lithium describes itself as an optimization mod that is not meant to change game mechanics (source 4), and Purpur as a drop-in replacement for Paper (source 5).
- One shared worldThe unmodified Vanilla server generated a world once with seed 20261121, containing 31,329 fully generated chunks. Each run started from a fresh copy of this world.
- Separate coresThe server was pinned to cores 0 to 3 and the load generator to cores 4 to 6. The measurement script was not pinned to any core. During the measurement, the raw data was packed every 15 minutes at lowest priority on core 7. The entry messung: 7 in plan.json was not implemented as a core pinning. Server and RCON were reachable only on the loopback address, and online-mode=false was set only in this isolated setup. Erik Friedl approved the Minecraft EULA for these test servers.
- Protocol botsA custom lightweight client speaks the real network protocol of version 26.3: login, configuration, play phase, keep-alive and teleport confirmations. The client sends chunk acknowledgements when it detects a chunk-batch message. This was the case for Vanilla and Fabric; for Paper and Purpur, it counted no such message in any run. The protocol database used by mineflayer did not yet include version 26.3 at the time of measurement. Each bot has a fixed seed, sends at a target rate of 20 position packets per second and moves at walking speed (0.2158 blocks per tick) between random waypoints. Every six bots form a group, and the groups are 256 blocks apart. The bots fly in creative mode at height 120.
- Levels and measurement windowsThe levels were 20, 30 and 40 bots, then in steps of five. After joining came a 60-second settling period and a 120-second measurement window with one sample every 5 seconds. A run ended after two levels above the limit or below 15 TPS.
- MetricsFor all programs, MSPT p50, p95 and p99 came from the /tick query command via RCON, which evaluates the last 100 ticks. The level value is the mean of these blocks over the measurement window. TPS was calculated from the progress of game time. In addition came RAM of the Java process, CPU per core group, CPU steal and the number of bots confirmed via the player list.
- Preliminary runs before the planBefore the plan was fixed, two pilot runs (Paper, Vanilla) and short trial runs of all combinations were carried out. After that, the level grid and measurement times were fixed; the 50 ms limit had been set beforehand. The preliminary runs are not included in the evaluation.
- ValidityFixed in advance: CPU steal below 2%, load generator below 70% CPU and below 20 ms timer lag (p99), complete samples, all bots connected as long as the server is not overloaded, and two sample bots demonstrably moving within the measurement window. Invalid runs would have been repeated; there were none.
- Capacity and differencesFixed in advance: the capacity of a run is the highest level up to which all levels stay below 50 ms p95 MSPT on average over the measurement window and all bots are connected. The capacity of a combination is the smallest value of its three runs. A difference counts only if the ranges of the repetitions do not overlap.
- Supplement added afterwardsAfter the measurement it became apparent that the tick rate in some cases already fell below 20 TPS while the p95 tick time was still below 50 ms on average. For this reason, the capacity at at least 19.5 TPS and the metrics at 50 bots were additionally evaluated. Level 50 is the highest at which all combinations still reached at least 19.5 TPS on average over the measurement window. These evaluations were not fixed in advance and are labeled accordingly in the data.
Measurement setup
- Game version
- Minecraft 26.3 (protocol 777)
- Vanilla
- 26.3
- Paper
- 26.3, build 257
- Purpur
- 26.3, build 2660
- Fabric
- Loader 0.20.1, installer 1.1.2
- Lithium
- 0.26.4 for 26.3
- Java
- Temurin 25.0.5.1+2
- Heap
- 10 GB (Xms equal to Xmx)
- JVM profile default G1
- -XX:+UseG1GC
- JVM profile Aikar flags
- Flags as per the PaperMC documentation
- JVM profile Generational ZGC
- -XX:+UseZGC
- Virtual machine
- 8 vCPU under KVM, 31.3 GiB RAM, 100 GiB virtual disk, Debian 12
- Reported CPU
- QEMU Virtual CPU, AuthenticAMD; the physical model cannot be determined from within the VM
- Cores
- Server 0 to 3, bots 4 to 6, backup 7, measurement script unpinned
- View distance
- 10 chunks, simulation distance 10 chunks
- Game mode
- Creative, difficulty normal
- Measurement period
- Nov 21, 2026, 20:12 UTC to Nov 22, 2026, 06:59 UTC
Sample and period
- Period
- Nov 21, 2026 to Nov 22, 2026
- Scope
- 18 Measurement runs
- Evaluated
- 18 Measurement runs
- Selection
- Six combinations of program and JVM profile with three repetitions each, in a randomized order fixed in advance.
Limits of what the study can show
- Bots are not playersThe bots fly, do not build, do not fight, do not chat, use no inventory and trigger no redstone circuits. Real players generate a different load. The absolute bot numbers can therefore not be transferred to a real server. The world was generated in advance and the bots did not leave this area. The load from generating new chunks is therefore not included.
- One VMMeasurements were taken on a single virtual machine with four cores for the server. On other hardware the values will differ, and the gaps between the programs may change as well.
- One test profileThere was one world, one seed, one movement profile and, apart from Lithium, no mods or plugins. The programs were not individually tuned. With an adjusted configuration or a different load, the order may look different. The day-night cycle ran along. Because all runs follow the same schedule, certain levels always fall into the night. The influence was not recorded separately in this measurement.
- Different default configurationThe default configuration of Paper and Purpur contains the setting entity-activation-range (animals and monsters 32 blocks), which is missing from the Vanilla and Fabric templates. Whether all programs performed the same simulation work was not checked, and entity counts were not recorded.
- MSPT and TPS divergeFor Vanilla, the p95 tick time at 55 bots was below 50 ms on average over the measurement window, but the tick rate was 16.31 to 17.16 TPS. The cause was not investigated. For Vanilla, the capacity fixed in advance is therefore 5 bots above the value at at least 19.5 TPS. For the five other combinations, both minimum values agree.
- Means instead of individual valuesThe limit of 50 ms applies to the mean of the 100-tick blocks over the measurement window. In 13 of 18 runs, at least one individual block at the capacity level was at 50 ms or above, at most 54.2 ms. The limit of 19.5 TPS also applies to the mean of the measurement window. In the levels of the capacity at at least 19.5 TPS, individual 5-second samples were below it, the lowest at 17.62 TPS (Vanilla, 50 bots).
- Drift over the measurement nightAt the same bot count, the p95 tick time rose with position in the run order, at 50 bots by roughly 0.11 ms per run (evaluated afterwards). The cause was not investigated in this measurement. The order was random but not balanced in blocks, so the spread of the repetitions also contains this effect.
- Metrics added afterwardsThe capacity at at least 19.5 TPS, the comparison level with 50 bots and the p99 comparison between Aikar flags and default G1 were chosen only after seeing the data. They are supplementary observations and not tests fixed in advance.
- Coarse grid, three repetitionsIn the relevant range the levels are five bots apart, and smaller differences cannot be measured. Three repetitions per combination show the spread only roughly. No statistical test was calculated.
- RAM values not comparableThe Aikar flags include AlwaysPreTouch, which reserves memory already at startup (source 1). The measured memory therefore depends on the JVM setting and does not show the program's requirements. This study makes no statement about which program needs less RAM.
- JVM comparison on Paper onlyThe three JVM profiles were measured only with Paper, a 10 GB heap and Java 25. The program comparison ran only with the Aikar flags. Neither result can readily be transferred to other combinations.
- SnapshotThe results apply to the stated versions and builds. Later versions may behave differently.
Data and evidence
To repeat the measurement you need a Linux machine with eight cores and about 32 GB RAM, as used here, plus Node.js and Python 3. The scripts expect the working directory /root/studien/haertetest, download Java and the same program versions as long as those downloads remain available, verify the checksums, generate the world and run the measurement plan. Seed, order, levels and rules are in plan.json. Different hardware, different builds or a different load on the host system lead to different absolute values.
| File | Size | SHA-256 |
|---|---|---|
| daten.csvMeasurements per level (CSV) | 35.6 KB | a2c2bfdfcc8ddfce2f6eac4044c15aebed57f4d0fab95b8a48af72c47e2f9e9e |
| daten.jsonMeasurements per level (JSON) | 94.6 KB | f798d400aba35617cf30f8c0cd2ad07cf6aace6e19e02300a0a1b822c6ac980e |
| laeufe.csvResult per run | 2.7 KB | 3a7f0af91987e982f8e95f97c60ef8330ea9ba8a9f4688e76c64f8514b3486d1 |
| zusammenfassung.csvResult per combination | 1.1 KB | 8ba8269c39e20d767c2a8c87dfb805fd0d937bb59bd448680f23401f0e66bd88 |
| vergleiche.csvPairwise comparisons | 4.7 KB | 99575f03f6a3f2075bcd20426a18e86d6d14ad5410618ebb101050c3fabcda9d |
| p99-aikar-g1.csvp99 Aikar flags and default G1 | 0.5 KB | 4d0b6c017068a81642cddc339b1623f271d5d543abaf0ea9e4722152907e64ac |
| proben.csvIndividual samples | 517.9 KB | 7fa08cf87e94c1e58c680ae9cec668ba6d56da228bda6be08849d8232db27c33 |
| plan.jsonMeasurement plan | 5.9 KB | cc335b804f52374f6c77afda5c55dd19dbf5fa03d595c8f3a9bb72cedd569192 |
| messprotokoll.txtMeasurement log | 8.1 KB | f234fa1d0a16b312777dd0b93fb12f0eb13161fbdbb13125501e034837705a64 |
| herkunft.txtOrigin of the programs | 1.4 KB | 5f9e921b1c4fe697d2e676b7b5777eeeae784a4c4418b211230e7ac333bb676d |
| liesmich.txtReadme | 1.4 KB | 8a444fae54933207d6f896459ac23511c541a656d0fc2da2a26b2f69091f0935 |
| 01-einrichten.shSetup | 3.4 KB | 27627241f4e7964be26e9b3ac2b3c1de74482e323348569f6e7ba6e9290b7ea9 |
| 02-vorlagen.pyWorld and templates | 4.1 KB | fc70d06d68b18d70425e0987794f06da257659424a168433c2048577258d4cd5 |
| 03-plan.pyMeasurement plan | 3.5 KB | 773d40d0d4dd1323c748a8c383ecf1ffb0fb91299cbea545f21144dfa083a034 |
| gemeinsam.pyShared components | 5.3 KB | da16b51f5b78fd55c070c1eaa9b43485cd7fd1654ce8c47e71df87b26bcb7e20 |
| messlauf.pyMeasurement run | 10.9 KB | 395ec326a334855d03a333879da16f900730d0307f06065f078f79af8de21637 |
| hauptlauf.pyMain measurement | 4.6 KB | 7b1602088aba30734e706e35ab894d4c918a91eec8b0846a68a2ee392c6cec2f |
| bots.mjsLoad generator | 14.3 KB | 6fa44c58b3c56a31fb77fb19602cc06aab93f08e8e8984a40bb95f2ad1064a10 |
| welt-pruefen.pyWorld check | 1.5 KB | e0bb7dcc1e2327e9b916355b3e3dba552751a2f6c13989ffd07f10c234c0940e |
| auswertung.mjsEvaluation | 10.0 KB | 42dc3452a2061bad7ede2d84610cdf8c8ef3adb5e7c6c7b8593c502aa2b85a33 |
| baue-studie.mjsStudy build | 39.2 KB | c0e76feab583afa11baf3bbce105203f4a3ebcf57e66adf44380bb7854a8b90b |
All checksums as a text file. Verify with sha256sum -c pruefsummen.txt.
Fields of the data files (72)
lauf- Identifier of the measurement run with position in the random order
kombination- Identifier made up of program and JVM profile
anzeige- Readable name of the combination
programm- Server program
jvm- JVM profile
wiederholung- Number of the repetition, 1 to 3
versuch- Attempt within the repetition, 1 is the first
reihenfolge- Position in the fixed random order
gueltig- Run meets all validity rules fixed in advance
lauf_gueltig- The associated run is valid
spielerBots- Target number of bots in this level
spieler_minBots- Smallest bot count confirmed via the player list in the measurement window
spieler_verbundenBots- Bot count confirmed via the player list in this sample
unter_grenze- p95 MSPT below 50 ms on average over the measurement window and all bots connected
p50_msms- Median of the tick time; in levels, mean of the 100-tick blocks
p95_msms- 95th percentile of the tick time; in levels, mean of the 100-tick blocks
p95_max_msms- Highest p95 value of a 100-tick block in the measurement window
p99_msms- 99th percentile of the tick time; in levels, mean of the 100-tick blocks
mspt_mittel_msms- Mean tick time according to /tick query
tpsTicks/s- Actual tick rate from the progress of game time; in levels, mean of the measurement window
tps_minTicks/s- Lowest tick rate of a 5-second sample in the measurement window
ram_gbGiB- Memory used by the Java process (RSS)
cpu_server_pct%- Mean utilization of the four server cores
cpu_bots_pct%- Mean utilization of the three cores of the load generator
steal_pct%- CPU steal across all eight cores
bots_verzug_p99_msms- Highest p99 lag of the timers in the load generator
getrenntBots- Bots unexpectedly disconnected up to this level
proben- Number of 5-second samples in the measurement window
zeit_unixs- Time of the sample as Unix time
erste_ueberlastBots- First level whose mean p95 MSPT is above the limit of 50 ms
kapazitaetBots- Capacity fixed in advance; in the summary, smallest value of the three runs
tps_bei_kapazitaetTicks/s- Mean tick rate at the capacity level
kapazitaet_tpsBots- Highest level up to which all levels held at least 19.5 TPS on average over the measurement window; in the summary, smallest value
kapazitaet_tps_190Bots- Like kapazitaet_tps, with 19.0 TPS as the limit
kapazitaet_tps_198Bots- Like kapazitaet_tps, with 19.8 TPS as the limit
p95_bei_50_msms- p95 MSPT at the level with 50 bots
p99_bei_50_msms- p99 MSPT at the level with 50 bots
tps_bei_50Ticks/s- Tick rate at the level with 50 bots
stufen_gemessen- Number of measured levels of the run
ungueltig_grund- Reason for invalid runs, otherwise empty
im_programmvergleich- Combination belongs to the program comparison
im_jvm_vergleich- Combination belongs to the comparison of the JVM profiles
laeufe_gueltig- Valid runs of the combination
laeufe_ungueltig- Invalid runs of the combination
kapazitaet_medianBots- Median of the capacity across the runs
kapazitaet_maxBots- Largest capacity value of the runs
kapazitaet_tps_medianBots- Median of kapazitaet_tps
kapazitaet_tps_maxBots- Largest value of kapazitaet_tps
tps_bei_kapazitaet_minTicks/s- Lowest tick rate at the capacity across the runs
p95_bei_50_min_msms- Smallest p95 value at 50 bots
p95_bei_50_max_msms- Largest p95 value at 50 bots
p99_bei_50_min_msms- Smallest p99 value at 50 bots
p99_bei_50_max_msms- Largest p99 value at 50 bots
ram_bei_50_gbGiB- Mean RSS at 50 bots
cpu_bei_50_pct%- Mean utilization of the server cores at 50 bots
vergleich- Program comparison or JVM comparison
kennzahl- Compared metric
regel- Whether the metric was fixed in advance or added afterwards
a- First combination
a_min- Smallest value of the first combination
a_max- Largest value of the first combination
b- Second combination
b_min- Smallest value of the second combination
b_max- Largest value of the second combination
unterschied_belegt- The ranges of the repetitions do not overlap
besser- Combination with the more favorable value, only for a demonstrated difference
aikar_minms- Smallest p99 value with Aikar flags at the level
aikar_maxms- Largest p99 value with Aikar flags at the level
g1_minms- Smallest p99 value with default G1 at the level
g1_maxms- Largest p99 value with default G1 at the level
aikar_niedriger- All three Aikar values are below all three G1 values
g1_niedriger- All three G1 values are below all three Aikar values
Sources
- [1]PaperMC (2026): Aikar's flags. https://docs.papermc.io/paper/aikars-flags/ Retrieved on Nov 21, 2026
- [2]Minecraft Wiki (2026): Commands/tick. https://minecraft.wiki/w/Commands/tick Retrieved on Nov 21, 2026
- [3]OpenJDK: JEP 490: ZGC: Remove the Non-Generational Mode. https://openjdk.org/jeps/490 Retrieved on Nov 21, 2026
- [4]Lithium-Projekt auf Modrinth (2026): Lithium. https://modrinth.com/mod/lithium Retrieved on Nov 21, 2026
- [5]PurpurMC: Purpur Documentation. https://purpurmc.org/docs/purpur/ Retrieved on Nov 21, 2026
Disclosure
Own interest
Nexthosting sells game servers itself, including servers for Minecraft, and therefore has a commercial interest in this topic. None of the tested programs comes from Nexthosting. The measurement ran on a test VM ordered by Nexthosting. The study does not rate any hosting providers' offers. Nexthosting is responsible for this study, contact: [email protected].
Authorship
Erik Friedl, Nexthosting. The publisher is Nexthosting.
License
This study and its data are available under CC BY 4.0.
How to cite this study
APA
Friedl, E. & Nexthosting. (2026). Server stress test: How many protocol bots can Vanilla, Paper, Purpur and Fabric with Lithium handle? (Version 1.0.0) [Report]. Nexthosting. https://doi.org/10.5281/zenodo.24440087
BibTeX
Download .bib@techreport{nexthosting2026serverhaertetestminecraft,
author = {Erik Friedl and {Nexthosting}},
title = {Server stress test: How many protocol bots can Vanilla, Paper, Purpur and Fabric with Lithium handle?},
institution= {Nexthosting},
type = {Study},
year = {2026},
version = {1.0.0},
url = {https://nexthosting.net/en/studies/server-haertetest-minecraft},
urldate = {2026-11-22},
language = {english},
doi = {10.5281/zenodo.24440087}
}
TY - RPRT AU - Erik Friedl AU - Nexthosting TI - Server stress test: How many protocol bots can Vanilla, Paper, Purpur and Fabric with Lithium handle? PY - 2026 DA - 2026/11/22 PB - Nexthosting ET - 1.0.0 LA - en DO - 10.5281/zenodo.24440087 UR - https://nexthosting.net/en/studies/server-haertetest-minecraft AB - 18 runs with protocol bots on a test VM: Paper 80, Purpur 75, Fabric with Lithium 65, Vanilla 55 (at 19.5 TPS: 50). On Paper, Generational ZGC came in below G1. KW - Minecraft server KW - Paper KW - Purpur KW - Fabric KW - Lithium KW - Vanilla KW - MSPT KW - TPS KW - Aikar flags KW - G1 KW - ZGC KW - Load test ER -
Changes and corrections
- 1.0.0Version
First version.