Benchmark Report
Tick spikes · Mob AI
The worst ticks are several times the mean. Average TPS stays fine while players feel the stutter.
- mspt_avg = 125.2 ms
- mspt_p99 = 267.97 ms
- spike_ratio = 2.14 ratio
- sample_count = 95 count
- spike_ticks = 8 count
- spike_window_ticks = 204 count
- spike_threshold_ms = 217.7 ms
- spike_ticks_with_gc = 0 count
- spike_ticks_with_throttle = 0 count
- spike_ticks_with_steal = 1 count
- spike_tick_ms = 2,594.3 ms
- spike_gc_pause_ms = 0 ms
- spike_throttled_ms = 0 ms
- spike_sample_interval_ms = 40 ms
- explained_share_threshold = 0.5 ratio
- explained_time_share_threshold = 0.25 ratio
Weakest test: Mob AI, 9.11 TPS on average. The 15 other tests hold at least 19.89 TPS.
Compared with other servers
Every measured server counts here, including private, anonymous and expired reports. Each server counts once, with its best run.
This report scores better than
80%
of the 6 servers running Paper & forks on Minecraft 26.3
Total score 122,369 · group median 104,129
Top quarter- Single Core
Better than 60% · 169,749 vs median 155,057
- Gameplay
Better than 60% · 68,541 vs median 65,583
- Exploration
Better than 80% · 96,275 vs median 80,634 · Top quarter
- Entities
Better than 60% · 77,742 vs median 74,573
- Mechanics
Better than 0% · 44,326 vs median 46,352 · Bottom quarter
- Hardware
Better than 80% · 54,463 vs median 48,218 · Top quarter
- Memory
Better than 83% · 33,800 vs median 30,056 · Top quarter
- Disk
Better than 60% · 59,404 vs median 56,257
- Multi-Core
Better than 60% · 91,154 vs median 84,128
Test by test
Each tick test's average MSPT (lower is better) and TPS, against the same group, each server with the same run as above. A test shows when at least 5 servers of the group passed it. The bar follows the MSPT.
- Block Physics
Better than 80% · MSPT 7.6 ms vs median 10 ms · TPS 20 vs median 20 · Top quarter
- Bone Meal Growth
Better than 80% · MSPT 0.8 ms vs median 1.1 ms · TPS 20 vs median 20 · Top quarter
- Chunk Loading
Better than 80% · MSPT 1.7 ms vs median 4.2 ms · TPS 19.9 vs median 19.7 · Top quarter
- Chunk Ticking
Better than 80% · MSPT 10.7 ms vs median 15 ms · TPS 20 vs median 20 · Top quarter
- Explosion
Better than 80% · MSPT 2.3 ms vs median 3.3 ms · TPS 19.9 vs median 19.9 · Top quarter
- Hopper
Better than 80% · MSPT 1.1 ms vs median 1.9 ms · TPS 20 vs median 20 · Top quarter
- Lighting Update
Better than 40% · MSPT 1.5 ms vs median 1.5 ms · TPS 20 vs median 20
- Mob AI
Better than 60% · MSPT 125.2 ms vs median 145 ms · TPS 9.1 vs median 8.2
- Redstone
Better than 60% · MSPT 20 ms vs median 22 ms · TPS 20 vs median 17.6
- Ticking Tile Entities
Better than 83% · MSPT 1.4 ms vs median 2.4 ms · TPS 20 vs median 20 · Top quarter
- World Save
Better than 60% · MSPT 0.7 ms vs median 0.9 ms · TPS 20 vs median 20
Scored with version 8.3.0 of the scoring rules. Groups of fewer than 5 servers are never used. Updated Oct 29, 2026, 12:40 AM UTC. Community trends · All measurements by CPU
What matters
Read from the measurements, never estimatedTick spikes
The worst ticks are several times the mean. Average TPS stays fine while players feel the stutter.
- mspt_avg = 125.2 ms
- mspt_p99 = 267.97 ms
- spike_ratio = 2.14 ratio
- sample_count = 95 count
- spike_ticks = 8 count
- spike_window_ticks = 204 count
- spike_threshold_ms = 217.7 ms
- spike_ticks_with_gc = 0 count
- spike_ticks_with_throttle = 0 count
- spike_ticks_with_steal = 1 count
- spike_tick_ms = 2,594.3 ms
- spike_gc_pause_ms = 0 ms
- spike_throttled_ms = 0 ms
- spike_sample_interval_ms = 40 ms
- explained_share_threshold = 0.5 ratio
- explained_time_share_threshold = 0.25 ratio
Non-default server configuration
Compare with care against a server running default settings.
- view-distance 6 (default 10)
- simulation-distance 4 (default 10)
›+1 more finding
The run started on a JVM that had just started, or that had not run a bench since it started: code was still being compiled and the heap still growing. The first run after a restart can differ from the next ones.
- jvm_uptime_s = 80 s
- bench_runs_since_start = 0 count
- threshold_s = 300 s
A finding stays silent rather than attribute a cause without evidence. An absent cause is an answer, not an omission.
Test by test
Problems first — select a row for the full test- Mob AIEntities9.11UnstableCritical
- Chunk LoadingExploration19.89ModerateUnstable
- ExplosionMechanics19.94StableUnstable
- Multi-Core CPUCPU19.99StableHealthy
- Block PhysicsSingle-Core Stress20.00StableHealthy
- Bone Meal GrowthMechanics20.00StableHealthy
- Chunk TickingExploration20.00StableHealthy
- Disk PerformanceMemory & Storage20.00StableHealthy
- HopperMechanics20.00StableHealthy
- Lighting UpdateExploration20.00ModerateHealthy
- Memory PerformanceMemory & Storage20.00StableHealthy
- Network (NBT)Network20.00StableHealthy
- RedstoneSingle-Core Stress20.00effective 20.0UnstableHealthy
- Single-Core (tick-aligned)CPU20.00StableHealthy
- Ticking Tile EntitiesMechanics20.00StableHealthy
- World SaveMechanics20.00StableHealthy
Settings your measurements justify
Each setting below is triggered by a measurement from this report, and names the test to replay to find out whether it helped.
Shrink the monster activation range
world-settings.default.entity-activation-range.monsters: 24Measured : mobAI.mspt.avg 125.2 ms (threshold 35.0 ms)
Mob AI costs more per tick than the budget allows. The activation range decides how far from a player a monster keeps thinking; shrinking it puts distant ones to sleep without removing them.
Test to replay : mobAI
Lower the monster spawn limit
spawn-limits.monsters: 50Currently : 70
Measured : mobAI.mspt.avg 125.2 ms (threshold 50.0 ms)
Mob AI is past the critical cost per tick, where shrinking the activation range is no longer enough. The spawn limit caps how many monsters exist per player at all, which is the only setting here that reduces the work rather than deferring it.
Test to replay : mobAI
These settings trade a little gameplay fidelity for tick budget. Apply one at a time and replay the named test: the effect stays a hypothesis until you have measured it.
Briefing for a language model
Paste this into the assistant of your choice to have it rewritten in prose. It already carries every number, so the model has none left to invent.
{
"format": "voxelbench.tuning.v1",
"instruction": "Rewrite each recommendation below as plain prose for a server operator. Every number you are allowed to use is already in this document: the measured value, its threshold, and the proposed setting. Do not compute, estimate or predict any other number, and never state an expected gain. Keep every setting name and value exactly as written.",
"server": {
"type": "paper",
"minecraft": "26.3",
"java": "25.0.4.1"
},
"recommendations": [
{
"id": "spigot.portee-activation-monstres",
"family": "spigot",
"measure": "mobAI.mspt.avg",
"value": 125.2,
"unit": "ms",
"threshold": 35,
"control_test": "mobAI",
"file": "spigot.yml",
"setting": "world-settings.default.entity-activation-range.monsters",
"proposed_value": "24",
"current_value": null
},
{
"id": "bukkit.limite-monstres",
"family": "bukkit",
"measure": "mobAI.mspt.avg",
"value": 125.2,
"unit": "ms",
"threshold": 50,
"control_test": "mobAI",
"file": "bukkit.yml",
"setting": "spawn-limits.monsters",
"proposed_value": "50",
"current_value": "70"
}
]
}Java start-up line
The options this server was started with, checked against its Java version. Corrections fix what is certainly wrong; leads to test are settings to measure, never applied for you.
Current line (20 arguments)
Values of system properties, agents and file paths are never shown: they appear as <redacted>.
-Xms15536M -Xmx15536M -XX:+UseG1GC -XX:+ParallelRefProcEnabled -XX:MaxGCPauseMillis=200 -XX:+UnlockExperimentalVMOptions -XX:+DisableExplicitGC -XX:+AlwaysPreTouch -XX:G1NewSizePercent=30 -XX:G1MaxNewSizePercent=40 -XX:G1HeapRegionSize=8M -XX:G1ReservePercent=20 -XX:G1HeapWastePercent=5 -XX:G1MixedGCCountTarget=4 -XX:InitiatingHeapOccupancyPercent=15 -XX:G1MixedGCLiveThresholdPercent=90 -XX:G1RSetUpdatingPauseTimePercent=5 -XX:SurvivorRatio=32 -XX:+PerfDisableSharedMem -XX:MaxTenuringThreshold=1
To correct
- DangerHeap too close to the machine's memory
-Xmx15536MThe heap may use 15536 MiB of the machine's 17922 MiB, which leaves 2386 MiB for everything else (code, threads, network buffers, the system). Under load the server can be killed without any Java error. The corrected line keeps 2689 MiB outside the heap.
Corrected line to copy
- Changed
-Xmx15536M-Xmx15233Mleaves 2689 MiB outside the heap. - Changed
-Xms15536M-Xms15233Mthe initial heap cannot exceed the maximum.
java -Xms15233M -Xmx15233M -XX:+UseG1GC -XX:+ParallelRefProcEnabled -XX:MaxGCPauseMillis=200 -XX:+UnlockExperimentalVMOptions -XX:+DisableExplicitGC -XX:+AlwaysPreTouch -XX:G1NewSizePercent=30 -XX:G1MaxNewSizePercent=40 -XX:G1HeapRegionSize=8M -XX:G1ReservePercent=20 -XX:G1HeapWastePercent=5 -XX:G1MixedGCCountTarget=4 -XX:InitiatingHeapOccupancyPercent=15 -XX:G1MixedGCLiveThresholdPercent=90 -XX:G1RSetUpdatingPauseTimePercent=5 -XX:SurvivorRatio=32 -XX:+PerfDisableSharedMem -XX:MaxTenuringThreshold=1 -jar server.jar noguiReplace server.jar with your server's jar. Paste the line where your server is started: the start-up parameters (or JVM arguments) of your hosting panel, or your start.sh / start.bat. Only the corrections above differ from your current line.
To test
Performance settings to try one at a time, then measure with a benchmark before and after. None of them is applied in the corrected line, and none comes with a promised gain.
A collector with very short pauses for large heaps, at the cost of more CPU and memory. It replaces the current collector: remove the current collector option and its own tuning options, then measure. From Java 23 the generational mode is the default.
-XX:+UseZGCNo measurement of this run points to it: it stays a lead to measure.
sourceRe-run and compare
- Apply the corrected line where your server starts (panel start-up parameters, start.sh).
- Restart the server.
- Run a new benchmark:
/bench start - Compare the new report with this one.
This report is the before; add the new report to the comparison once it is in. A single run on each side says little: the gap between two runs of an unchanged server is noise.
Run conditions
What the plugin saw of the host during the benchmark, and what the site concludes from it. The plugin reports facts; the verdicts and their thresholds are the site's, the same ones the diagnosis and the API use.
The reference computation timed during the run kept its pace, and the hypervisor took no significant CPU time.
A CPU quota, a reduced CPU set or a memory limit applies to the server.
Contention during the run
- Reference computation slowdown
- 0 % (threshold 15 %)3 timings: 12.6 ms · 12.5 ms · 12.5 ms (a slow first timing alone does not count) — judged on the median of at least 40 samples per timing (plugin 2.6.2); host noise 3.5 %, the threshold widens to it, from 15 to 30 % — wait for the server to settle before Hopper, after chunk loading: 0.0 s (limit 30.0 s, after chunk generation)
- CPU time stolen by the hypervisor
- 0 % (threshold 5 %)
- Scheduling periods ended throttled by the quota
- 0 % (threshold 25 %)0 ms
- Time a task waited for a core
- 0.1 % (threshold 40 %)
Limits seen by the server
- CPU quota
- 8 cores
- CPU set
- 0-7 (8 cores)
- Memory limit
- 17.8 GiB
- Host logical CPUs
- 8
- Host memory
- 17.5 GiB
- Container signals
- dockerenv, cgroup-v2-namespace, overlay-root, pterodactyl-env
Storage
- Filesystem
- xfs · Local filesystem
- Direct 4K reads
- Honoured (70 µs against 0.3 µs from cache)
Single-core measurement
- Throughput (wall clock)
- 652.9M ops/s652.7M ops/s over the thread's CPU time
- Thread CPU time / wall time
- 1.000 (threshold 0.900)Below 1, the thread was not running for part of its slices (quota, neighbours, hypervisor steal when counted, JVM pauses). Close to 1 proves nothing.
- Spread from second to second
- 1.7 % (threshold 40 %)p10 652.5M · p50 659.6M · p90 664M ops/s over 9 seconds
- GC pauses during the measurement
- 0 ms (threshold 50 ms)0 collections, 0 ms of collection time
- Highest current CPU frequency
- not exposed by this systemadvertised maximum: 3,187 MHz
Witnesses of the single-core test (plugin 2.5.0). The score does not read this test, and none of these figures changes the run's stability verdict.
Host, memory and server state
- Time a task waited for storage
- 5.3 % (all tasks at once: 5.3 %)includes the disk test, which waits for storage by design: not judged for the run
- Time a task stalled on memory
- 0 % (all tasks at once: 0 %) (threshold 5 %)
- CPU quota during the run
- 8 coresno change observed
- Container memory
- peak 16,993 MiBlimit events: high 0 · max 0 · OOM kills 0
- Runnable tasks on the machine
- up to 7.5 on average during a test, for 8 logical CPUs
- Host uptime at the start
- 18.1 h
- Server at the start of the run
- JVM up for 1.3 minfirst bench since the JVM startedoutside the bench world: 0 players · 254 chunks · 137 entities
Run pressures from plugin 2.4.0, the rest from plugin 2.6.0. None of these figures changes the score or the run's stability verdict; the diagnosis names the tests they concern.
Thresholds and their reasons are published in the benchmark methodology. Rules rc-7.
Every measurement
Each test in full. Tests in trouble are open; the others fold to their summary line.
Single-Core Stress
RedstoneHealthy20.00 TPS20.0 ms34.9s
Single-core stress via intensive redstone circuit updates.
- Heavy garbage collection5 collections, 39ms total
Percentiles and stability
Raw data (68 fields)
Block PhysicsHealthy20.00 TPS7.6 ms24.9s
Single-core stress via falling-block physics.
- Heavy garbage collection3 collections, 35ms total
Percentiles and stability
Raw data (71 fields)
Exploration
Chunk LoadingUnstable19.89 TPS1.7 ms7.7s
Chunk generation and loading speed across dispersed zones.
- Heavy garbage collection1 collections, 244ms total
- The run did not hold steadyTPS P1=15.00<15
Percentiles and stability
Raw data (59 fields)
Chunk TickingHealthy20.00 TPS10.7 ms40.9s
Server performance when ticking many loaded chunks with random tick updates.
- Heavy garbage collection2 collections, 84ms total
Percentiles and stability
Raw data (63 fields)
Lighting UpdateHealthy20.00 TPS1.5 ms24.6s
Lighting recomputation throughput on block place / break.
Percentiles and stability
Raw data (56 fields)
Entities
Mob AICritical9.11 TPS125.2 ms30.3s
Complex mob AI behaviours (villagers, iron golems, hostile mobs).
- Heavy garbage collection3 collections, 253ms total
- The run did not hold steadyTPS avg=9.11<19, TPS P1=5.00<15, TPS StdDev=2.44>2, TPS min=5.00 (55% of avg)
Percentiles and stability
Raw data (80 fields)
Mechanics
HopperHealthy20.00 TPS1.1 ms19.2s
Intensive hopper chain item transfers.
- Heavy garbage collection1 collections, 79ms total
Percentiles and stability
Raw data (72 fields)
ExplosionUnstable19.94 TPS2.3 ms17.0s
Massive TNT explosions and block destruction.
- The run did not hold steadyTPS P1=14.99<15
Percentiles and stability
Raw data (110 fields)
Ticking Tile EntitiesHealthy20.00 TPS1.4 ms35.9s
Many ticking tile entities (furnaces, hoppers, etc.).
Percentiles and stability
Raw data (58 fields)
World SaveHealthy20.00 TPS0.7 ms4.8s
World save speed and TPS impact during save operations.
Percentiles and stability
Raw data (56 fields)
Bone Meal GrowthHealthy20.00 TPS0.8 ms23.3s
Rapid crop and tree growth via bone meal.
Percentiles and stability
Raw data (60 fields)
CPU
Single-Core (tick-aligned)Healthy20.00 TPS2.8 ms10.0s
Single-thread compute throughput measured on the main thread under the tick scheduler.
Percentiles and stability
Raw data (61 fields)
Multi-Core CPUHealthy19.99 TPS0.5 ms10.0s
Parallel computation throughput across worker threads.
Percentiles and stability
Raw data (56 fields)
Memory & Storage
Memory PerformanceHealthy20.00 TPS0.8 ms19.6s
RAM bandwidth and latency.
JVM measurements — do not reflect actual bare-metal performance.
Percentiles and stability
Raw data (56 fields)
Disk PerformanceHealthy20.00 TPS0.7 ms49.7s
Sequential and random 4K read/write performance.
JVM measurements — do not reflect actual bare-metal performance.
Percentiles and stability
Raw data (55 fields)
Network
Network (NBT)Healthy20.00 TPS16.2 ms20.0s
Serialisation throughput — proxy for the rate at which the server can ship world data to clients.
- Heavy garbage collection5 collections, 74ms total
Percentiles and stability
Raw data (61 fields)
›Machine and serverIntel(R) Core(TM) i9-14900K · Paper · 26.3
Score Breakdown
Detailed sub-category scores