Forge and NeoForge hosting for heavy modpacks

What fills the heap is not the mod jar, it is what the mod registers on startup: blocks, items, fluids, recipes, tags, entities and classes generated at runtime. One tech mod with thousands of recipes weighs more than ten decoration mods put together.

The other source is the world. Every dimension a mod adds carries its own chunks and its own entities for as long as somebody is in there. The memory graph has steps, not a ramp: it rises at boot, settles, and rises again when someone opens a dimension or starts a factory that holds chunks loaded. Sizing by the first day figure is what jams the server in week two.

Most of the startup is not disk. It is bytecode transformation, when Mixin and coremods rewrite classes across dozens of mods, plus building the registries and reading the recipes, tags and loot tables of the pack. Spawn generation depends on storage and clock, but it only happens on the first start.

Two conclusions follow. Do not judge a host by the first boot, compare the second one. And if the time suddenly grows between restarts, the suspect is a new mod or a datapack that reloads everything, not the hardware.

There is a floor and a slope. The floor is the fixed cost of the pack, which sits in memory even with the server empty. The slope is the cost per player, in practice the cost per loaded chunk. On a heavy pack the floor dominates, and ten people ask for more memory than thirty on a plain survival.

Measuring is quick: restart, note the empty heap, run a peak and note it again. That difference divided by the players online gives the slope of your pack. Correcting upward is cheap, 12GB at R$ 132,00 against 16GB at R$ 176,00. What costs is skipping three steps without measuring anything.

The server closes a tick every 50 milliseconds to hold 20 TPS. A JVM pause longer than that is a lost tick, and a few hundred milliseconds turn into the freeze everyone feels at the same moment. A bigger heap does not remove pauses, it changes their nature: fewer collections, each one with more objects to walk through.

That is why doubling Xmx almost never solves it. A heap far larger than the live data set gives the collector little reason to clean, until it cleans everything at once, and the 24GB server stutters more than it did on 12GB. Xms equal to Xmx, G1 aligned to the tick and a profiler at peak will say whether the bottleneck is heap or a mod.