snek-tech-blog/content/posts/jvm-gc-and-mc.md

---
title: "What is GC? How does it work in JVM?"
date: 2023-01-28T20:15:10+11:00
draft: false
showSummary: true
summary: "Tuning Garbage Collection on Minecraft servers is something of an arcane art that few people seem to understand
at a fundamental level. Let's remedy that. We'll start at fundamentals of GC, the JVM, and use that to form the
foundations of our GC choice and how we tune it."
series:
  - "JVM, GC, and Minecraft"
series_order: 1
---

# Part One: What is GC? How does it work?

> **(Tammy)** To preempt any fears you may have, Ashe has assured us that this will be written at a level where someone
> who just wants to run their minecraft server well will understand.
>
> **(Ashe)** And for those of you who do have a technical understanding of Java, and/or sysadmin, I hope this deepens
> your understanding of the topics at hand.
>
> **(Doll)** Will this one be able to understand?
>
> **(Ashe)** That's the idea, yeah.
>
> **(Doll)** YAY!

## What is GC?

GC is an initialism for Garbage Collection, and refers to the process of a program (or operating system, or anything
else) freeing unused memory for re-use. We don't need to go super in-depth here,
except to say that the garbage collection alorithm and mechanism will determine when and how garbage collection is done,
and what limitations there are to GC.

## Why do I care?

The choice of garbage collection algorithm determines a lot about how the underlying program will behave.
Broadly, there are three categories of garbage collection algorithms:

- Stop-the-world
- Concurrent
- Hybrid

### Stop-the-world garbage collection

{{< hover "STW" >}} Stop-the-world {{< /hover >}} GC algorithms, as the name suggests, halt program execution while
freeing memory. Often, they will also compact live objects as there's no risk of a reference changing
while program execution is paused.

The Serial GC in JVM (which is the default in many implementations) is an example of a stop-the-world GC algorithm.

{{< hover "Stop-the-world" >}}STW{{< /hover >}} garbage collectors are ideal when short pauses in program execution
are acceptable and the running program should have the lion's share of resources.

### Concurrent garbage collection

Concurrent garbage collectors *do not* pause program execution while doing clean-up, and as such run on a thread
parallel to program execution. Concurrent algorithms cannot make the same assumptions as stop-the-world algorithms
about program execution and as such must go to much more effort to ensure that a reference is stale or dead before
cleaning it up.
As such, they tend to be much more CPU and RAM intensive in exchange for never forcing the program to stop.

Concurrent GC algorithms are ideal when program pauses are unacceptable (such as in realtime or networked applications),
and when some inefficiency in CPU and RAM usage is an acceptable price to pay for this.

The {{< hover "Z Garbage Collector" >}}ZGC{{< /hover >}} in Java is an example a fully concurrent garbage collector.

### Hybrid

Hybrid garbage collectors are, as the name suggests, some hybrid of the above.
They will do some of their work concurrently, but do require {{< hover "Stop-the-world" >}}STW{{< /hover >}}
pauses to perform some of their functions.

The Concurrent-Mark-and-Sweep collector (deprecated since JVM 9, removed in JVM 14) is an example of a hybrid GC.
It marks memory as stale concurrently using a small amount of the available threads,
and then stops-the-world to do collection and compaction.

The {{< hover "Garbage-First Garbage Collector" >}}G1GC{{< /hover >}}, which replaced the Concurrent-Mark-and-Sweep
collector, is another example.

### A secret fourth thing

> **(Tammy)** Wait. Could you... simply not?
>
> **(Ashe)** Good catch. Yes! It's even useful sometimes!

You could simply *not* collect garbage and allow the program to continue using more and more memory until
it either terminates, or inevitably runs out of resources and crashes.

> **(Lorelai)** You only need last long enough to fulfill your purpose. I can see the appeal.

Right. The appeal here is that you get the full processing power of the machine, *and* you never have to stop the world
to collect garbage; it's the best of both worlds. However, this also means that you're either relying on the programmer
to clean up their own garbage (in languages that allow this), or simply allowing the program to use more
and more memory until the host runs out of available memory and the program crashes.
For our purposes today, it is worth noting that disabling the garbage collector in the JVM means that
no garbage collection will be done, and, as such no memory will ever be freed,
since Java lacks explicit destruction of objects.

But Lorelai has the right idea. This is the approach used by some stock-trading software,
where any delay is unacceptable,
and the software has a bounded lifecycle. It needs to run for 8 hours every weekday. As such, so long as it doesn't crash
within those 8 hours, it can do whatever.

> **(Ashe)** As an aside, this is how we came across the phrase "Giant mainframes with 256GB of RAM running JVM 8,
crashing every 9 hours, all according to keikaku[^1]"

[^1]: Keikaku means plan.

## The JVM and Garbage Collection

To understand the particularities of the various GCs available in modern Java, we'll have to take a small digression into
how the Java Virtual Machine itself works. The Java Virtual Machine is precisely that: a virtual machine.
It has its own instruction set, memory management, everything. It is in most ways a full architecture[^2].
As such, the Java compiler compiles Java source code into JVM bytecode, which the virtual machine then executes.
Java achieves its ability to "run on anything, including toasters" by publishing
{{< hover "Java Runtime Environments" >}}JREs{{< /hover >}} for many different underlying architectures.

[^2]: As an interesting side note, there are even some SOCs that can run Java bytecode natively.

Java as a programming language lands squarely in the Object-Orientated camp,
but makes several critical diversions from C and friends to make itself easier to port to different architectures:

- No exposed pointers
- No explicit free/destroy

### No Exposed pointers

Java as a language refuses to expose raw pointers to the programmer. It does this to make porting JVM to
different architectures with potentially wildly different memory architectures much easier, as the programmer can never
refer to a pointer by address or perform pointer arithmetic. This means that the particular address of any particular
object can change at any time. This makes garbage-collection significantly easier, as it means that the GC algorithm
can copy live data at any time, which makes memory compaction much easier.

### No explicit free/destroy

The JVM does not provide any mechanism to explicitly free or destroy an object.
Even `System.GC()` is nothing more than a suggestion.

This is important as it takes a lot of control away from the programmer and grants it to the runtime instead.
The person running the program can determine the GC algorthim, how often it runs, what its runtime targets are,
and so on, based on the particular performance characteristics that they need. One could argue that this could
equally be done by the programmer or designer of the program, but in doing this, JVM again gains a lot of portability,
as it means that the sole arbiter of when an object is freed is the runtime.

## How does this relate to Minecraft?

> **(Doll)** But Miss Ashe! How does this help us get all the dolls on one biiiig Minecraft Server?

I'm glad you asked!

An interesting note here is that *because* the JVM abstracts away memory management from the programmer, it is up to
the deployment to choose the correct GC and parameters to ensure that their program performs the way that it is
intended to. Many programs ship with relatively sensible defaults, however Minecraft (and especially modded Minecraft)
was notoriously bad about this. There was some folk wisdom back in the day the suggested using
the {{< hover "Concurrent-Mark-and-Sweep" >}}CMS{{< /hover >}} garbage collector along side some sensible,
but unexplained, parameters that seemed to perform well, but this forum post seems to either have been lost to bit-rot,
or is too well-hidden for our google-fu.

In more modern times, various launchers have converged on some relatively sane defaults, using CMS at first
(as it performed best with Minecraft's workload), and moving to G1GC more recently, as Minecraft itself moved to using
JDK versions where CMS was deprecated (or even gone!).

PolyMC currently ships the following on all of their servers:
```bash
# Xmx and Xms set the maximum and minimum RAM usage, respectively.
# They can take any number, followed by an M or a G.
# M means Megabyte, G means Gigabyte.
# For example, to set the maximum to 3GB: -Xmx3G
# To set the minimum to 2.5GB: -Xms2500M
# A good default for a modded server is 4GB.

-Xms4G
-Xmx6G
-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
```

If this is all Greek to you, don't worry! We'll go over this in our next post and see where we can go from here.

Suffice it to say there's plenty of room for improvement here without degrading performance, especially if your
workload differs significantly from what these defaults assume.

## Next Time!

This is probably a good place to call it for today. While we haven't gotten into actually tuning Minecraft GC,
I think this forms a good groundwork for understanding why we're going to make certain decisions down the line.

So we've gone over how and why the JVM manages memory, so next time, we'll focus on the various GCs that modern JREs
ship with, and what their various advantages and disadvantages are. If you'd like a head-start,
the [Oracle GC-Tuning Documentation](https://docs.oracle.com/en/java/javase/17/gctuning/preface.html)
is extremely thorough!

> **(Tammy)** Okay, I managed to keep up with all of that, nice work.
>
> **(Tess)** Was there ever any doubt?