Repository navigation
Releases
Releases · tsurutanmen/minecraft-server-parallelism
v1.1.0: patches for the clustered case
Adds section 5: two small server patches for the case where players cluster, measured with the same bot swarm (4,800 villagers, 16 groups 256 blocks apart).
- Remember empty POI searches (Paper and Folia): skips villager workstation/bed searches that are proven to find nothing until a POI changes or the villager moves more than 8 blocks. Clustered Folia region 0.24-0.26 → 2.81-2.82 TPS; Paper ver/26.1.2 4.1-4.4 → 6.9 TPS. In a job-taking test, villagers still took a newly placed composter in 6-12 s, the same as unpatched servers.
- Per-thread cache of Folia's ownership check: removes the concurrent-table lookups from the profile but did not reduce tick time. Kept as a negative result.
Fix: Folia prints MSPT with a thousands separator, and v1.0.0's results/SUMMARY.md read only the digits after the comma (clustered Folia showed 989 / 727 ms instead of 4,110 / 3,727 ms). TPS values and Paper/ShreddedPaper numbers were not affected.
Diffs are in patches/, raw data in results/opt_*.
v1.0.0: first set of measurements
First release of the measurements and scripts (Minecraft Java Edition 26.1.2, one PC).
- Chunk generation: vanilla vs C2ME vs C2ME + ScalableLux vs C2ME + OpenCL, with a block-level terrain comparison against vanilla
- Mob pushing and pathfinding on CPU (numba) vs GPU (CUDA via CuPy)
- Paper vs Folia vs ShreddedPaper under a 32-bot swarm, players spread out and clustered
- JFR breakdown of Paper's main thread and the Amdahl bound for splitting work by kind
Raw data, including invalid runs, is in results/. See README.md for the findings and limitations.