Smooth Server
面向 Fabric 的服务器性能模组整合包。
Smooth Server 汇集了适用于 Minecraft Java 1.17.1 至 26.3(含其间稳定补丁版本)的服务器性能模组。它针对刻处理、内存使用、红石、地形生成和网络通信进行了优化。请将其安装在你的 Fabric 服务器上;玩家客户端无需安装此整合包。
性能测试数据
我们将未修改的已发布 26.3 Fabric-SmoothS v1 整合包与 无模组的 Fabric 26.3 环境在 六项合成工作负载下进行了对比,每项负载重复五次配对测试,共计 60 次服务器运行测量。排除 Fixture 验证试点阶段。每次试验均从其工作负载基准构建的世界副本开始;配对内的配置顺序交替进行,且同一时间仅运行一个服务器。
刻计算与新生地形
| 工作负载 |
无模组 Fabric |
Smooth Server |
使用该整合包节省的时间 |
配对结果 |
| 混合控制:400 只拥有 AI 的独立圈养僵尸 + 1,024 个空漏斗 |
2.010 ms/tick |
0.810 ms/tick |
59.7% |
全部 5 对均有改善 |
| 仅 AI:400 只独立圈养僵尸 |
1.610 ms/tick |
0.760 ms/tick |
52.8% |
全部 5 对均有改善 |
| 闲置漏斗:4,096 个未通电的空漏斗 |
0.390 ms/tick |
0.020 ms/tick |
94.9% |
全部 5 对均有改善 |
| 活跃传输:双箱之间 256 条独立的四漏斗通道 |
0.130 ms/tick |
0.100 ms/tick |
23.1% |
全部 5 对均有改善 |
| 红石:16 组时钟驱动的 13 x 13 红石粉网格 |
1.600 ms/tick |
0.110 ms/tick |
93.1% |
全部 5 对均有改善 |
| 新生地形:八批每批 64 个新目标区块,共 512 个 |
13.969 s |
13.423 s |
3.9% |
混合 / 结论不明 |
数值为 每种配置五次运行的中位数;百分比基于这些中位数进行比较。时间越短越好。刻行数据测量的是经过 4,000 次预热刻和 60 秒正常节奏后,在 12,000 次刻冲刺期间的计算耗时。该冲刺移除了正常的 20 TPS 上限;其吞吐量不代表生产环境的 TPS 或玩家承载量。控制台冲刺计时四舍五入至 0.01 ms,这限制了极低开销工作负载的精度。
所有五项主要的配对测量在每项刻工作负载中均有改善。这描述了这些试验中的 一致性,而非统计置信度声明。地形测试包含一组退化的配对,因此其中位数时间的减少 并非一致或已确立的地形加速效果。地形的 CPU 使用中位数也增加了 4.6%,且配对结果不一。
正常节奏与权衡
在正常的 20 TPS 目标下,中位数 滚动平均刻时间如下:
| 工作负载 |
无模组 Fabric |
Smooth Server |
使用该整合包节省的时间 |
| 混合控制 |
4.2 ms/tick |
2.2 ms/tick |
47.6% |
| 仅 AI |
3.6 ms/tick |
2.1 ms/tick |
41.7% |
| 闲置漏斗 |
1.2 ms/tick |
0.2 ms/tick |
83.3% |
| 活跃传输 |
0.6 ms/tick |
0.5 ms/tick |
16.7% |
| 红石粉 |
1.9 ms/tick |
0.4 ms/tick |
78.9% |
这些是控制台 tick query 命令的滚动平均值,而非单个刻的百分位数。在正常节奏下,活跃传输的配对结果不一;其他四项工作负载在所有配对中均有改善。
该整合包并未提升所有指标。 在中位数常驻 Java 进程内存方面,五项刻工作负载的平均值 高出 8.8-15.3%,且在所有六项工作负载中启动时间 延长了 28.6-32.8%;这两项测量的所有配对均出现退化。中位数正常节奏下的 CPU 使用率在闲置漏斗工作中从 0.013 增加至 0.041 核心等效值(+207.4%),在活跃传输工作中从 0.026 增加至 0.048(+84.2%),所有配对均出现退化。这些百分比是基于较小的绝对 CPU 使用量计算的。混合控制、AI 和红石的正常节奏 CPU 结果不一。在混合控制、AI、闲置漏斗和红石的所有配对中,冲刺 CPU 时间均减少;活跃传输冲刺 CPU 结果不一,且两种配置的中位数相等。
测试条件与范围
测试使用 Windows 10 Pro、Temurin Java 25.0.4、Fabric Loader 0.19.5、最大堆内存 2 GiB 以及对 JVM 可见的四处理器,并设置 4 GiB 进程内存限制和 50% 主机 CPU 上限。测试期间 没有玩家;自然刷怪、随机刻、昼夜循环和天气循环均已禁用。磁盘缓存已预热,JVM 预热过程有界;这些是在共享主机上进行的本地测试。
Fixture 检查覆盖了实体数量、活跃漏斗进度和项目守恒,在测量期间 stocked sources 持续存在,在正常节奏和冲刺期间保持红石活动,并完成新生地形。初始世界哈希值在工作负载内匹配,已安装的模组文件和提供的配置与已发布的工件进行了核对。未添加任何测量模组。
地形模拟已冻结,生成预热在单独的非测量区域进行,批次之间释放了强制加载的区块。其计时涵盖 地形加载和准备,包括可能的相邻区块生成。内存测量描述的是 Java 进程的工作集,而不仅仅是堆内存。
收益归因于 整个整合包,而非单独的模组。结果仅适用于 Minecraft 26.3 上的这些合成工作负载;它们不能预测每个世界、机器、玩家负载或其他受支持版本的表现。未测量 FPS 和网络吞吐量,也不保证服务器完全无延迟。基准测试套件参考:20261011T195733Z-23636。
整合包内容
可用模组因 Minecraft 版本而异。26.3 构建版包含:
- Lithium 和 ServerCore 用于服务器刻优化。
- FerriteCore 和 ModernFix-mVUS 用于内存及其他性能提升。
- Fast Noise 和 Structure Layout Optimizer 用于地形与结构生成。
- Alternate Current 用于红石粉优化,Krypton 用于网络优化。
- Chunky 用于可选的区块预生成,spark 用于诊断性能问题。
- TT20、No Chat Reports、Fabric API、Resourceful Config 以及所需的依赖项 ZConfig。
仅选择兼容的稳定版模组发布版本。缺失的模组会在每个版本的更新日志中列出;当有兼容的稳定版本可用时,会重新考虑加入。26.3 构建版本缺失了 Async、C2ME、Immersive Optimization 和 ScalableLux。
以原版为核心,存在技术差异
本整合包不添加任何新内容或进度系统。部分已包含的模组会影响技术行为:Alternate Current 会改变红石更新顺序,TT20 会在低 TPS 情况下改变时间逻辑,包括流体、随机刻和方块破坏。包含 Immersive Optimization 的版本还会减少某些实体的刻数。No Chat Reports 会更改聊天报告行为。在使用对数据敏感的红石农场或建筑时应考虑这些因素;本整合包并不能保证与原版的完全一致。
选择版本
请选择与你的 Minecraft 精确版本号 匹配的文件。Fabric 是当前自动发布版本所支持的加载器。
整合包的频道标识表示其 18 个项目参考目录中有多少是可用的:
- Release(正式版): 缺少零到两个参考项目。
- Beta(测试版): 缺少三个到八个。
- Alpha(早期版): 缺少九个或以上。
这些标签描述的是整合包的完整度;已安装的模组发布版本仍然是稳定的。请参阅更新日志以了解实际的缺失内容和变更。
安装
使用支持 Modrinth 整合包格式的服务器工具安装 .mrpack 文件,并使用该 Minecraft 版本所需的 Java 版本。在更新时请保留世界文件和配置文件的备份。区块预生成和分析是可选工具;它们不会作为整合包基准测试的一部分自动运行。
Smooth Server
Server performance mods, packaged for Fabric.
Smooth Server brings together server performance mods for Minecraft Java 1.17.1 through 26.3, including the stable patch releases in between. It targets tick processing, memory use, redstone, terrain generation and networking. Install it on your Fabric server; players do not need this modpack on their clients.
Measured performance
We compared the unchanged published 26.3 Fabric-SmoothS v1 pack against Fabric 26.3 without mods across six synthetic workloads, with five paired repetitions each: 60 measured server runs. Fixture-validation pilots were excluded. Each trial started from a fresh copy of its workload's baseline-built world; configuration order alternated within pairs, with one server running at a time.
Tick computation and fresh terrain
| Workload |
Fabric without mods |
Smooth Server |
Less time with the pack |
Paired outcome |
| Mixed control: 400 individually penned zombies with AI + 1,024 empty hoppers |
2.010 ms/tick |
0.810 ms/tick |
59.7% |
All 5 pairs improved |
| AI only: 400 individually penned zombies |
1.610 ms/tick |
0.760 ms/tick |
52.8% |
All 5 pairs improved |
| Idle hoppers: 4,096 empty, unpowered hoppers |
0.390 ms/tick |
0.020 ms/tick |
94.9% |
All 5 pairs improved |
| Active transfers: 256 independent four-hopper lanes between double chests |
0.130 ms/tick |
0.100 ms/tick |
23.1% |
All 5 pairs improved |
| Redstone: 16 clocked 13 x 13 dust grids |
1.600 ms/tick |
0.110 ms/tick |
93.1% |
All 5 pairs improved |
| Fresh terrain: eight batches of 64 new target chunks, 512 total |
13.969 s |
13.423 s |
3.9% |
Mixed / inconclusive |
Values are medians of five runs per configuration; percentages compare those medians. Lower time is better. Tick rows measure computation during a 12,000-tick sprint after 4,000 warmup ticks and 60 seconds at normal cadence. The sprint removes the normal 20 TPS cap; its throughput does not establish production TPS or player capacity. Console sprint timings are rounded to 0.01 ms, which limits precision for very cheap workloads.
All five paired primary measurements improved in each tick workload. This describes consistency in these trials, not a statistical confidence claim. Terrain included a regressing pair, so its smaller median time is not a consistent or established terrain speedup. Terrain CPU time also increased by 4.6% in the median, with mixed paired outcomes.
Normal cadence and tradeoffs
At the normal 20 TPS target, median rolling average tick times were:
| Workload |
Fabric without mods |
Smooth Server |
Less time with the pack |
| Mixed control |
4.2 ms/tick |
2.2 ms/tick |
47.6% |
| AI only |
3.6 ms/tick |
2.1 ms/tick |
41.7% |
| Idle hoppers |
1.2 ms/tick |
0.2 ms/tick |
83.3% |
| Active transfers |
0.6 ms/tick |
0.5 ms/tick |
16.7% |
| Redstone dust |
1.9 ms/tick |
0.4 ms/tick |
78.9% |
These are console tick query rolling averages, not individual-tick percentiles. Active-transfer paired results were mixed at normal cadence; the other four workloads improved in every pair.
The pack did not improve every metric. Median resident Java process memory was 8.8-15.3% higher across the five tick workloads, and startup took 28.6-32.8% longer across all six workloads; every pair regressed on those measurements. Median normal-cadence CPU use increased from 0.013 to 0.041 core equivalents for idle hoppers (+207.4%) and from 0.026 to 0.048 for active transfers (+84.2%), with regressions in every pair. Those percentages start from small absolute CPU usage. Normal-cadence CPU results were mixed for mixed control, AI and redstone. Sprint CPU time decreased for mixed control, AI, idle hoppers and redstone in every pair; active-transfer sprint CPU results were mixed, with equal configuration medians.
Test conditions and scope
Tests used Windows 10 Pro, Temurin Java 25.0.4, Fabric Loader 0.19.5, a 2 GiB maximum heap and four processors visible to the JVM, with a 4 GiB process memory limit and 50% host CPU cap. There were no players; natural spawning, random ticks, day and weather cycles were disabled. Disk caches were warm and the JVM warmup was bounded; these were local tests on a shared host.
Fixture checks covered entity counts, active-hopper progress and item conservation, stocked sources throughout measurement, sustained redstone activity during normal cadence and sprint, and fresh terrain completion. Initial world hashes matched within each workload, and installed mod files and supplied configuration were checked against the published artifact. No measurement mods were added.
Terrain simulation was frozen, generation was warmed in a separate unmeasured region, and forced chunks were released between batches. Its timing covers terrain loading and preparation, including possible neighboring generation. Memory measurements describe the Java process working set, not just its heap.
Gains are attributed to the full pack, not isolated mods. Results apply to these synthetic workloads on Minecraft 26.3; they do not predict every world, machine, player load or other supported version. FPS and network throughput were not measured, and a lag-free server is not guaranteed. Benchmark suite reference: 20261011T195733Z-23636.
What the pack includes
Available mods vary with the Minecraft version. The 26.3 build includes:
- Lithium and ServerCore for server tick optimizations.
- FerriteCore and ModernFix-mVUS for memory and other performance improvements.
- Fast Noise and Structure Layout Optimizer for terrain and structure generation.
- Alternate Current for redstone dust optimizations, and Krypton for networking optimizations.
- Chunky for optional chunk pregeneration, and spark for diagnosing performance.
- TT20, No Chat Reports, Fabric API, Resourceful Config and the required ZConfig dependency.
Only compatible stable mod releases are selected. Missing mods are listed in each version's changelog; they are reconsidered when a compatible stable release becomes available. The 26.3 build is missing Async, C2ME, Immersive Optimization and ScalableLux.
Vanilla-focused, with technical differences
The pack adds no new content or progression systems. Some included mods affect technical behavior: Alternate Current changes redstone update order, and TT20 changes timing under low TPS, including fluids, random ticks and block breaking. Versions that include Immersive Optimization also reduce some entity ticking. No Chat Reports changes chat-reporting behavior. These choices should be considered when using sensitive farms or redstone builds; the pack is not a guarantee of identical vanilla behavior.
Choosing a version
Pick the file matching your exact Minecraft version. Fabric is the supported loader for the current automated releases.
The pack's channels indicate how much of its 18-project reference catalog is available:
- Release: zero to two reference projects missing.
- Beta: three to eight missing.
- Alpha: nine or more missing.
These labels describe pack completeness; installed mod releases are still stable. See the changelog for the actual omissions and changes.
Installation
Install the .mrpack with a server tool that supports Modrinth packs, using the Java version required by that Minecraft release. Keep a backup of your world and configuration when updating. Chunk pregeneration and profiling are optional tools; they do not run automatically as part of the pack's benchmark.
评论 0
登录后参与评分
还没有评论,来抢沙发吧