_content/doc/gc-guide.html

<!--{
	"Title": "A Guide to the Go Garbage Collector",
	"Path":  "/doc/gc-guide",
  	"Breadcrumb": true
}-->

<!--
NOTE: In this document and others in this directory, the convention is to
set fixed-width phrases with non-fixed-width spaces, as in
<code>hello</code> <code>world</code>.
Do not send CLs removing the interior tags from such phrases.
-->

<style>
.gc-guide-graph {
  display: inline-block;
  position: relative;
  width: 100%;
  vertical-align: top;
  overflow: hidden;
}

.gc-guide-graph-controls {
  display: flex;
  flex-direction: row;
  flex-wrap: wrap;
  justify-content: space-around;
  align-items: center;
  width: 100%;
}

.gc-guide-graph-controls div {
  display: flex;
  flex-direction: row;
  flex-wrap: nowrap;
  align-items: center;
  padding-left: 5px;
  padding-right: 5px;
  height: 24px;
  width: auto;
}

.gc-guide-equation {
  display: block;
  text-align: center;
}
</style>

<h2 id="Introduction">Introduction</h2>

<p>
This guide is intended to aid advanced Go users in better understanding their
application costs by providing insights into the Go garbage collector.
It also provides guidance on how Go users may use these insights to improve
their applications' resource utilization.
It does not assume any knowledge of garbage collection, but does assume
familiarity with the Go programming language.
</p>

<p>
The Go language takes responsibility for arranging the storage of Go values;
in most cases, a Go developer need not care about where these values are stored,
or why, if at all.
In practice, however, these values often need to be stored in computer
<b>physical memory</b> and physical memory is a finite resource.
Because it is finite, memory must be managed carefully and recycled in order to
avoid running out of it while executing a Go program.
It's the job of a Go implementation to allocate and recycle memory as needed.
</p>

<p>
Another term for automatically recycling memory is <b>garbage collection</b>.
At a high level, a garbage <i>collector</i> (or GC, for short) is a system that
recycles memory on behalf of the application by identifying which parts of memory
are no longer needed.
The Go standard toolchain provides a runtime library that ships with every
application, and this runtime library includes a garbage collector.
</p>

<p>
Note that the existence of a garbage collector as described by this guide
is not guaranteed by the <a href="/ref/spec">Go specification</a>, only that
the underlying storage for Go values is managed by the language itself.
This omission is intentional and enables the use of radically different
memory management techniques.
</p>

<p>
Therefore, this guide is about a specific implementation of the Go programming
language and <i>may not apply to other implementations</i>.
Specifically, this following guide applies to the standard toolchain (the
<code>gc</code> Go compiler and tools).
Gccgo and Gollvm both use a very similar GC implementation so many of the
same concepts apply, but details may vary.
</p>

<p>
Furthermore, this is a living document and will change over time to best
reflect the latest release of Go.
This document currently describes the garbage collector as of Go 1.19.
</p>


<h3 id="Where_Go_Values_Live">Where Go Values Live</h3>

<p>
Before we dive into the GC, let's first discuss the memory that doesn't need to
be managed by the GC.
</p>

<p>
For instance, non-pointer Go values stored in local variables will likely not be
managed by the Go GC at all, and Go will instead arrange for memory to be
allocated that's tied to the
<a href="https://go.dev/ref/spec#Declarations_and_scope">lexical scope</a> in
which it's created.
In general, this is more efficient than relying on the GC, because the Go
compiler is able to predetermine when that memory may be freed and emit
machine instructions that clean up.
Typically, we refer to allocating memory for Go values this way as
"stack allocation," because the space is stored on the goroutine stack.
</p>

<p>
Go values whose memory cannot be allocated this way, because the Go compiler
cannot determine its lifetime, are said to <i>escape to the heap</i>.
"The heap" can be thought of as a catch-all for memory allocation, for when Go
values need to be placed <i>somewhere</i>.
The act of allocating memory on the heap is typically referred to as "dynamic
memory allocation" because both the compiler and the runtime can make very few
assumptions as to how this memory is used and when it can be cleaned up.
That's where a GC comes in: it's a system that specifically identifies and
cleans up dynamic memory allocations.
</p>

<p>
There are many reasons why a Go value might need to escape to the heap.
One reason could be that its size is dynamically determined.
Consider for instance the backing array of a slice whose initial size is
determined by a variable, rather than a constant.
Note that escaping to the heap must also be transitive: if a reference to a
Go value is written into another Go value that has already been determined to
escape, that value must also escape.
</p>

<p>
Whether a Go value escapes or not is a function of the context in which it is
used and the Go compiler's escape analysis algorithm.
It would be fragile and difficult to try to enumerate precisely when values
escape: the algorithm itself is fairly sophisticated and changes between Go
releases.
For more details on how to identify which values escape and which do not, see
the section on
<a href="#Eliminating_heap_allocations">eliminating heap allocations</a>.
</p>

<h3 id="Tracing_Garbage_Collection">Tracing Garbage Collection</h3>

<p>
Garbage collection may refer to many different methods of automatically
recycling memory; for example, reference counting.
In the context of this document, garbage collection refers to <b>tracing</b>
garbage collection, which identifies in-use, so-called <b>live</b>, objects by
following pointers transitively.
</p>

<p>
Let's define these terms more rigorously.
</p>

<ul>
	<li>
		<p>
		<b>Object</b>&mdash;An object is a dynamically allocated piece of memory
		that contains one or more Go values.
		</p>
	</li>
	<li>
		<p>
		<b>Pointer</b>&mdash;A memory address that references any value within an
		object.
		This naturally includes Go values of the form <code>*T</code>, but also includes
		parts of built-in Go values.
		Strings, slices, channels, maps, and interface values all contain memory
		addresses that the GC must trace.
		</p>
	</li>
</ul>

<p>
Together, objects and pointers to other objects form the <b>object graph</b>.
To identify live memory, the GC walks the object graph starting at the
program's <b>roots</b>, pointers that identify objects that are definitely
in-use by the program.
Two examples of roots are local variables and global variables.
The process of walking the object graph is referred to as <b>scanning</b>.
</p>

<p>
This basic algorithm is common to all tracing GCs.
Where tracing GCs differ is what they do once they discover memory is live.
Go's GC uses the mark-sweep technique, which means that in order to keep track
of its progress, the GC also <b>marks</b> the values it encounters as live.
Once tracing is complete, the GC then walks over all memory in the heap and
makes all memory that is <i>not</i> marked available for allocation.
This process is called <b>sweeping</b>.
</p>

<p>
One alternative technique you may be familiar with is to actually <i>move</i>
the objects to a new part of memory and leave behind a forwarding pointer that
is later used to update all the application's pointers.
We call a GC that moves objects in this way a <b>moving</b> GC; Go has a
<b>non-moving</b> GC.
</p>

<h2 id="The_GC_cycle">The GC cycle</h2>

<p>
Because the Go GC is a mark-sweep GC, it broadly operates in two phases: the
mark phase, and the sweep phase.
While this statement might seem tautological, it contains an important insight:
it's not possible to release memory back to be allocated until <i>all</i> memory
has been traced, because there may still be an un-scanned pointer keeping
an object alive.
As a result, the act of sweeping must be entirely separated from the act of
marking.
Furthermore, the GC may also not be active at all, when there's no GC-related
work to do.
The GC continuously rotates through these three phases of sweeping, off, and
marking in what's known as the <b>GC cycle</b>.
For the purposes of this document, consider the GC cycle starting with sweeping,
turning off, then marking.
</p>

<p>
The next few sections will focus on building intuition for the costs of the
GC to aid users in tweaking GC parameters for their own benefit.
</p>

<h3 id="Understanding_costs">Understanding costs</h3>

<p>
The GC is inherently a complex piece of software built on even more complex
systems.
It's easy to become mired in detail when trying to understand the GC and
tweak its behavior.
This section is intended to provide a framework for reasoning about the cost
of the Go GC and tuning parameters.
</p>

<p>
To begin with, consider this model of GC cost based on three simple axioms.
</p>

<ol>
	<li>
		<p>
		The GC involves only two resources: CPU time, and physical memory.
		</p>
	</li>
	<li>
		<p>
		The GC's memory costs consist of live heap memory, new heap memory
		allocated before the mark phase, and space for metadata that, even
		if proportional to the previous costs, are small in comparison.
		</p>

		<p>
		<i>
		Note: live heap memory is memory that was determined to be live by the
		previous GC cycle, while new heap memory is any memory allocated in
		the current cycle, which may or may not be live by the end.
		</i>
		</p>
	</li>
	<li>
		<p>
		The GC's CPU costs are modeled as a fixed cost per cycle, and a
		marginal cost that scales proportionally with the size of the live
		heap.
		</p>
		<p>
		<i>
		Note: Asymptotically speaking, sweeping scales worse than marking and
		scanning, as it must perform work proportional to the size of the whole
		heap, including memory that is determined to be not live (i.e. "dead").
		However, in the current implementation sweeping is so much faster than
		marking and scanning that its associated costs can be ignored in this
		discussion.
		</i>
		</p>
	</li>
</ol>

<p>
This model is simple but effective: it accurately categorizes the dominant
costs of the GC.
However, this model says nothing about the magnitude of these costs, nor
how they interact.
To model that, consider the following situation, referred to from here on as
the <b>steady-state</b>.
</p>

<ul>
	<li>
		<p>
		The rate at which the application allocates new memory (in bytes per
		second) is constant.
		</p>

		<p>
		<i>
		Note: it's important to understand that this allocation rate is
		completely separate from whether or not this new memory is live.
		None of it could be live, all of it could be live, or some of it could
		be live.
		(On top of this, some old heap memory could also die, so it's not
		necessarily the case that if that memory is live, the live heap size
		grows.)
		</i>
		</p>

		<p>
		<i>
		To put this more concretely, consider a web service that allocates
		2 MiB of total heap memory for each request that it handles.
		During the request, at most 512 KiB of that 2 MiB stays live while
		the request is in flight, and when the service is finished handling
		the request, all that memory dies.
		Now, for the sake of simplicity suppose each request takes about 1
		second to handle end-to-end.
		Then, a steady stream of requests, say 100 requests per second,
		results in an allocation rate of 200 MiB/s and a 50 MiB peak live
		heap.
		</i>
		</p>
	</li>
	<li>
		<p>
		The application's object graph looks roughly the same each time
		(objects are similarly sized, there's a roughly constant number of
		pointers, the maximum depth of the graph is roughly constant).
		</p>
		<p>
		Another way to think about this is that the marginal costs of
		GC are constant.
		</p>
	</li>
</ul>

<p>
<i>
Note: the steady-state may seem contrived, but it's representative of the
behavior of an application under some constant workload.
Naturally, workloads can change even while an application is executing, but
typically application behavior looks like a bunch of these steady-states strung
together with some transient behavior in between.
</i>
</p>

<p>
<i>
Note: the steady-state makes no assumptions about the live heap.
It may be growing with each subsequent GC cycle, it may shrink, or it may
stay the same.
However, trying to encompass all of these situations in the explanations
to follow is tedious and not very illustrative, so the guide will focus
on examples where the live heap remains constant.
The <a href="#GOGC">GOGC section</a> explores the non-constant live heap
scenario in some more detail.
</i>
</p>

<p>
In the steady-state while the live heap size is constant, every GC cycle is
going to look identical in the cost model as long as the GC executes after
the same amount of time has passed.
That's because in that fixed amount of time, with a fixed rate of allocation
by the application, a fixed amount of new heap memory will be allocated.
So with the live heap size constant, and that new heap memory constant,
memory use is always going to be the same.
And because the live heap is the same size, the marginal GC CPU costs will
be the same, and the fixed costs will be incurred at some regular interval.
</p>

<p>
Now consider if the GC were to shift the point at which it runs <i>later</i>
in time.
Then, more memory would be allocated but each GC cycle would still incur
the same CPU cost.
However over some other fixed window of time fewer GC cycles would finish,
resulting in a lower overall CPU cost.
The opposite would be true if the GC decided to start <i>earlier</i> in time:
less memory would be allocated and CPU costs would be incurred more often.
</p>

<p>
This situation represents the fundamental trade-off between CPU time and memory
that a GC can make, controlled by how often the GC actually executes.
In other words, the trade-off is entirely defined by <b>GC frequency</b>.
</p>

<p>
One more detail remains to be defined, and that's when the GC should decide
to start.
Note that this directly sets the GC frequency in any particular steady-state,
defining the trade-off.
In Go, deciding when the GC should start is the main parameter which the user
has control over.
</p>

<h3 id="GOGC">GOGC</h3>

<p>
At a high level, GOGC determines the trade-off between GC CPU and memory.
</p>

<p>
It works by determining the target heap size after each GC cycle, a target value
for the total heap size in the next cycle.
The GC's goal is to finish a collection cycle before the total heap size
exceeds the target heap size.
Total heap size is defined as the live heap size at the end of the previous
cycle, plus any new heap memory allocated by the application since the previous
cycle.
Meanwhile, target heap memory is defined as:
</p>

<p class="gc-guide-equation">
<i>
Target heap memory = Live heap + (Live heap + GC roots) * GOGC / 100
</i>
</p>

<p>
As an example, consider a Go program with a live heap size of 8 MiB, 1 MiB
of goroutine stacks, and 1 MiB of pointers in global variables.
Then, with a GOGC value of 100, the amount of new memory that will be allocated
before the next GC runs will be 10 MiB, or 100% of the 10 MiB of work, for a
total heap footprint of 18 MiB.
With a GOGC value of 50, then it'll be 50%, or 5 MiB.
With a GOGC value of 200, it'll be 200%, or 20 MiB.
</p>

<p>
<i>
Note: GOGC includes the root set only as of Go 1.18.
Previously, it would only count the live heap.
Often, the amount of memory in goroutine stacks is quite small and the live
heap size dominates all other sources of GC work, but in cases where programs
had hundreds of thousands of goroutines, the GC was making poor judgements.
</i>
</p>

<p>
The heap target controls GC frequency: the bigger the target, the longer the GC
can wait to start another mark phase and vice versa.
While the precise formula is useful for making estimates, it's best to think of
GOGC in terms of its fundamental purpose: a parameter that picks a point in the
GC CPU and memory trade-off.
The key takeaway is that <b>doubling GOGC will double heap memory overheads and
roughly halve GC CPU cost</b>, and vice versa.
(To see a full explanation as to why, see the
<a href="#Additional_notes_on_GOGC">appendix</a>.)
</p>

<p>
<i>
Note: the target heap size is just a target, and there are several reasons why
the GC cycle might not finish right at that target.
For one, a large enough heap allocation can simply exceed the target.
However, other reasons appear in GC implementations that go beyond the
<a href="#Understanding_costs">GC model</a> this guide has been using thus far.
For some more detail, see the <a href="#Latency">latency section</a>, but the
complete details may be found in the <a href="#Additional_resources">additional
resources</a>.
</i>
</p>

<p>
GOGC may be configured through either the <code>GOGC</code> environment
variable (which all Go programs recognize), or through the
<a href="https://pkg.go.dev/runtime/debug#SetGCPercent"><code>SetGCPercent</code></a>
API in the <code>runtime/debug</code> package.
</p>

<p>
Note that GOGC may also be used to turn off the GC entirely (provided the
<a href="#Memory_limit">memory limit</a> does not apply) by setting
<code>GOGC=off</code> or calling <code>SetGCPercent(-1)</code>.
Conceptually, this setting is equivalent to setting GOGC to a value of
infinity, as the amount of new memory before a GC is triggered is unbounded.
</p>

<p>
To better understand everything we've discussed so far, try out the interactive
visualization below that is built on the
<a href="#Understanding_costs">GC cost model</a> discussed earlier.
This visualization depicts the execution of some program whose non-GC work takes
10 seconds of CPU time to complete.
In the first second it performs some initialization step (growing its live heap)
before settling into a steady-state.
The application allocates 200 MiB in total, with 20 MiB live at a time.
It assumes that the only relevant GC work to complete comes from the live heap,
and that (unrealistically) the application uses no additional memory.
</p>

<p>
Use the slider to adjust the value of GOGC to see how the application responds
in terms of total duration and GC overhead.
Each GC cycle ends while the new heap drops to zero.
The time taken while the new heap drops to zero is the combined time for the
mark phase for cycle N, and the sweep phase for the cycle N+1.
Note that this visualization (and all the visualizations in this guide) assume
the application is paused while the GC executes, so GC CPU costs are fully
represented by the time it takes for new heap memory to drop to zero.
This is only to make visualization simpler; the same intuition still applies.
The X axis shifts to always show the full CPU-time duration of the program.
Notice that additional CPU time used by the GC increases the overall duration.
</p>

<div class="gc-guide-graph" data-workload='[
	{"duration": 1.0, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 1.00, "oldDeathRate": 0.00},
	{"duration": 9.0, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 0.00, "oldDeathRate": 0.00}
]' data-config='{
	"fixedCost": 0.04,
	"otherMem": 0,
	"GOGC": "graph1-gogc",
	"memoryLimit": 100000
}'></div>
<div class="gc-guide-graph-controls">
	<div>
		GOGC
		<input type="range" min="0" max="10" step="0.005" value="6.64" id="graph1-gogc">
		<div class="gc-guide-counter" id="graph1-gogc-display"></div>
	</div>
</div>

<p>
Notice that the GC always incurs some CPU and peak memory overhead.
As GOGC increases, CPU overhead decreases, but peak memory increases
proportionally to the live heap size.
As GOGC decreases, the peak memory requirement decreases at the expense of
additional CPU overhead.
</p>

<p>
<i>
Note: the graph displays CPU time, not wall-clock time to complete the program.
If the program runs on 1 CPU and fully utilizes its resources, then these are
equivalent.
A real-world program likely runs on a multi-core system and does not 100%
utilize the CPUs at all times.
In these cases the wall-time impact of the GC will be lower.
</i>
</p>

<p>
<i>
Note: the Go GC has a minimum total heap size of 4 MiB, so if the GOGC-set
target is ever below that, it gets rounded up.
The visualization reflects this detail.
</i>
</p>

<p>
Here's another example that's a little bit more dynamic and realistic.
Once again, the application takes 10 CPU-seconds to complete without the GC, but
the steady-state allocation rate increases dramatically half-way through, and
the live heap size shifts around a bit in the first phase.
This example demonstrates how the steady-state might look when the live heap
size is actually changing, and how a higher allocation rate leads to more
frequent GC cycles.
</p>

<div class="gc-guide-graph" data-workload='[
	{"duration": 1.0, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 1.00, "oldDeathRate": 0.00},
	{"duration": 1.0, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 0.00, "oldDeathRate": 0.50},
	{"duration": 1.0, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 0.50, "oldDeathRate": 0.00},
	{"duration": 1.0, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 0.00, "oldDeathRate": 0.50},
	{"duration": 1.0, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 0.50, "oldDeathRate": 0.00},
	{"duration": 5.0, "allocRate": 200, "scanRate": 1024, "newSurvivalRate": 0.02, "oldDeathRate": 1.00}
]' data-config='{
	"fixedCost": 0.04,
	"otherMem": 0,
	"GOGC": "graph2-gogc",
	"memoryLimit": 100000
}'></div>
<div class="gc-guide-graph-controls">
	<div>
		GOGC
		<input type="range" min="0" max="10" step="0.005" value="6.64" id="graph2-gogc">
		<div class="gc-guide-counter" id="graph2-gogc-display"></div>
	</div>
</div>

<h3 id="Memory_limit">Memory limit</h3>

<p>
Until Go 1.19, GOGC was the sole parameter that could be used to modify the GC's
behavior.
While it works great as a way to set a trade-off, it doesn't take into account
that available memory is finite.
Consider what happens when there's a transient spike in the live heap size:
because the GC will pick a total heap size proportional to that live heap size,
GOGC must be configured such for the <i>peak</i> live heap size, even if in the
usual case a higher GOGC value provides a better trade-off.
</p>

<p>
The visualization below demonstrates this transient heap spike situation.
</p>

<div class="gc-guide-graph" data-workload='[
	{"duration": 1.0, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 1.00, "oldDeathRate": 0.00},
	{"duration": 4.0, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 0.00, "oldDeathRate": 0.00},
	{"duration": 0.5, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 1.00, "oldDeathRate": 0.00},
	{"duration": 0.5, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 0.00, "oldDeathRate": 0.00},
	{"duration": 0.5, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 0.00, "oldDeathRate": 0.50},
	{"duration": 3.5, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 0.00, "oldDeathRate": 0.00}
]' data-config='{
	"fixedCost": 0.04,
	"otherMem": 0,
	"GOGC": "graph3-gogc",
	"memoryLimit": 100000
}'></div>
<div class="gc-guide-graph-controls">
	<div>
		GOGC
		<input type="range" min="0" max="10" step="0.005" value="6.64" id="graph3-gogc">
		<div class="gc-guide-counter" id="graph3-gogc-display"></div>
	</div>
</div>

<p>
If the example workload is running in a container with a bit over 60 MiB of
memory available, then GOGC can't be increased beyond 100, even though the rest
of the GC cycles have the available memory to make use of that extra memory.
Furthermore, in some applications, these transient peaks can be rare and hard to
predict, leading to occasional, unavoidable, and potentially costly
out-of-memory conditions.
</p>

<p>
That's why in the 1.19 release, Go added support for setting a runtime memory
limit.
The memory limit may be configured either via the <code>GOMEMLIMIT</code>
environment variable which all Go programs recognize, or through the
<code>SetMemoryLimit</code> function available in the <code>runtime/debug</code>
package.
</p>

<p>
This memory limit sets a maximum on the <i>total amount of memory that the Go
runtime can use</i>.
The specific set of memory included is defined in terms of
<a href="https://pkg.go.dev/runtime#MemStats"><code>runtime.MemStats</code></a>
as the expression
</p>

<p>
<code>Sys</code> <code>-</code> <code>HeapReleased</code>
</p>

<p>
or equivalently in terms of the
<a href="https://pkg.go.dev/runtime/metrics"><code>runtime/metrics</code></a>
package,
</p>

<p>
<code>/memory/classes/total:bytes</code> <code>-</code> <code>/memory/classes/heap/released:bytes</code>
</p>

<p>
Because the Go GC has explicit control over how much heap memory it uses, it
sets the total heap size based on this memory limit and how much other memory
the Go runtime uses.
</p>

<p>
The visualization below depicts the same single-phase steady-state workload from
the GOGC section, but this time with an extra 10 MiB of overhead from the Go
runtime and with an adjustable memory limit.
Try shifting around both GOGC and the memory limit and see what happens.
</p>

<div class="gc-guide-graph" data-workload='[
	{"duration": 1.0, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 1.00, "oldDeathRate": 0},
	{"duration": 9.0, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 0.00, "oldDeathRate": 0}
]' data-config='{
	"fixedCost": 0.04,
	"otherMem": 10,
	"GOGC": "graph4-gogc",
	"memoryLimit": "graph4-memlimit"
}'></div>
<div class="gc-guide-graph-controls">
	<div>
		GOGC
		<input type="range" min="0" max="10" step="0.005" value="6.64" id="graph4-gogc">
		<div class="gc-guide-counter" id="graph4-gogc-display"></div>
	</div>
	<div>
		Memory Limit
		<input type="range" min="1" max="100" step="0.5" value="100" id="graph4-memlimit">
		<div class="gc-guide-counter" id="graph4-memlimit-display"></div>
	</div>
</div>

<p>
Notice that when the memory limit is lowered below the peak memory that's
determined by GOGC (42 MiB for a GOGC of 100), the GC runs more frequently to
keep the peak memory within the limit.
</p>

<p>
Returning to our previous example of the transient heap spike, by setting a
memory limit and turning up GOGC, we can get the best of both worlds: no memory
limit breach, and better resource economy.
Try out the interactive visualization below.
</p>

<div class="gc-guide-graph" data-workload='[
	{"duration": 1.0, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 1.00, "oldDeathRate": 0.00},
	{"duration": 4.0, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 0.00, "oldDeathRate": 0.00},
	{"duration": 0.5, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 1.00, "oldDeathRate": 0.00},
	{"duration": 0.5, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 0.00, "oldDeathRate": 0.00},
	{"duration": 0.5, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 0.00, "oldDeathRate": 0.50},
	{"duration": 3.5, "allocRate": 20, "scanRate": 1024, "newSurvivalRate": 0.00, "oldDeathRate": 0.00}
]' data-config='{
	"fixedCost": 0.04,
	"otherMem": 0,
	"GOGC": "graph5-gogc",
	"memoryLimit": "graph5-memlimit"
}'></div>
<div class="gc-guide-graph-controls">
	<div>
		GOGC
		<input type="range" min="0" max="10" step="0.005" value="6.64" id="graph5-gogc">
		<div class="gc-guide-counter" id="graph5-gogc-display"></div>
	</div>
	<div>
		Memory Limit
		<input type="range" min="1" max="100" step="0.5" value="100" id="graph5-memlimit">
		<div class="gc-guide-counter" id="graph5-memlimit-display"></div>
	</div>
</div>

<p>
Notice that with some values of GOGC and the memory limit, peak memory use
stops at whatever the memory limit is, but that the rest of the program's
execution still obeys the total heap size rule set by GOGC.
</p>

<p>
This observation leads to another interesting detail: even when GOGC is set to
off, the memory limit is still respected!
In fact, this particular configuration represents a <i>maximization of resource
economy</i> because it sets the minimum GC frequency required to maintain some
memory limit.
In this case, <i>all</i> of the program's execution has the heap size rise to
meet the memory limit.
</p>

<p>
Now, while the memory limit is clearly a powerful tool, <b>the use of
a memory limit does not come without a cost</b>, and certainly doesn't
invalidate the utility of GOGC.
</p>

<p>
Consider what happens when the live heap grows large enough to bring total
memory use close to the memory limit.
In the steady-state visualization above, try turning GOGC off and then slowly
lowering the memory limit further and further to see what happens.
Notice that the total time the application takes will start to grow in an
unbounded manner as the GC is constantly executing to maintain an impossible
memory limit.
</p>

<p>
This situation, where the program fails to make reasonable progress due to
constant GC cycles, is called <b>thrashing</b>.
It's particularly dangerous because it effectively stalls the program.
Even worse, it can happen for exactly the same situation we were trying to
avoid with GOGC: a large enough transient heap spike can cause a program to
stall indefinitely!
Try reducing the memory limit (around 30 MiB or lower) in the transient heap
spike visualization and notice how the worst behavior specifically starts with
the heap spike.
</p>

<p>
In many cases, an indefinite stall is worse than an out-of-memory condition,
which tends to result in a much faster failure.
</p>

<p>
For this reason, the memory limit is defined to be <b>soft</b>.
The Go runtime makes no guarantees that it will maintain this memory limit
under all circumstances; it only promises some reasonable amount of effort.
This relaxation of the memory limit is critical to avoiding thrashing behavior,
because it gives the GC a way out: let memory use surpass the limit to avoid
spending too much time in the GC.
</p>

<p>
How this works internally is the GC sets an upper limit on the amount
of CPU time it can use over some time window (with some hysteresis for very
short transient spikes in CPU use).
This limit is currently set at roughly 50%, with a <code>2 * GOMAXPROCS</code>
CPU-second window.
The consequence of limiting GC CPU time is that the GC's work is delayed,
meanwhile the Go program may continue allocating new heap memory, even beyond
the memory limit.
</p>

<p>
The intuition behind the 50% GC CPU limit is based on the worst-case impact
on a program with ample available memory.
In the case of a misconfiguration of the memory limit, where it is set too
low mistakenly, the program will slow down at most by 2x, because the GC
can't take more than 50% of its CPU time away.
</p>

<p>
<i>
Note: the visualizations on this page do not simulate the GC CPU limit.
</i>
</p>

<h4 id="Suggested_uses">Suggested uses</h4>

<p>
While the memory limit is a powerful tool, and the Go runtime takes steps to
mitigate the worst behaviors from misuse, it's still important to use it
thoughtfully.
Below is a collection of tidbits of advice about where the memory limit is most
useful and applicable, and where it might cause more harm than good.
</p>

<ul>
	<li>
		<p>
		<b>Do</b> take advantage of the memory limit when the execution
		environment of your Go program is entirely within your control, and
		the Go program is the only program with access to some set of resources
		(i.e. some kind of memory reservation, like a container memory limit).
		</p>
		<p>
		A good example is the deployment of a web service into containers with
		a fixed amount of available memory.
		</p>
		<p>
		<b>In this case, a good rule of thumb is to leave an additional 5-10%
		of headroom to account for memory sources the Go runtime is unaware of.
		</b>
		</p>
	</li>
	<li>
		<p>
		<b>Do</b> feel free to adjust the memory limit in real time to adapt to
		changing conditions.
		</p>
		<p>
		A good example is a cgo program where C libraries temporarily need to
		use substantially more memory.
		</p>
	</li>
	<li>
		<p>
		<b>Don't</b> set GOGC to off with a memory limit if the Go program
		might share some of its limited memory with other programs, and those
		programs are generally decoupled from the Go program.
		Instead, keep the memory limit since it may help to curb undesirable
		transient behavior, but set GOGC to some smaller, reasonable value for
		the average case.
		</p>
		<p>
		While it may be tempting to try and "reserve" memory for co-tenant
		programs, unless the programs are fully synchronized (e.g. the Go
		program calls some subprocess and blocks while its callee executes),
		the result will be less reliable as inevitably both programs will
		need more memory.
		Letting the Go program use less memory when it doesn't need it will
		generate a more reliable result overall.
		This advice also applies to overcommit situations, where the sum of
		memory limits of containers running on one machine may exceed the
		actual physical memory available to the machine.
		</p>
	</li>
	<li>
		<p>
		<b>Don't</b> use the memory limit when deploying to an execution
		environment you don't control, especially when your program's memory
		use is proportional to its inputs.
		</p>
		<p>
		A good example is a CLI tool or a desktop application.
		Baking a memory limit into the program when it's unclear what kind of
		inputs it might be fed, or how much memory might be available on the
		system can lead to confusing crashes and poor performance.
		Plus, an advanced end-user can always set a memory limit if they wish.
		</p>
	</li>
	<li>
		<p>
		<b>Don't</b> set a memory limit to avoid out-of-memory conditions when
		a program is already close to its environment's memory limits.
		</p>
		<p>
		This effectively replaces an out-of-memory risk with a risk of
		severe application slowdown, which is often not a favorable trade,
		even with the efforts Go makes to mitigate thrashing.
		In such a case, it would be much more effective to either increase the
		environment's memory limits (and <i>then</i> potentially set a memory
		limit) or decrease GOGC (which provides a much cleaner trade-off than
		thrashing-mitigation does).
		</p>
	</li>
</ul>

<h3 id="Latency">Latency</h3>

<p>
The visualizations in this document have modeled the application as paused while
the GC is executing.
GC implementations do exist that behave this way, and they're referred to as
"stop-the-world" GCs.
</p>

<p>
The Go GC, however, is not fully stop-the-world and does most of its work
concurrently with the application.
This is primarily to reduce application <i>latencies</i>.
Specifically, the end-to-end duration of a single unit of computation (e.g. a
web request).
Thus far, this document mainly considered application <i>throughput</i> (e.g.
web requests handled per second).
Note that each example in the <a href="#The_GC_cycle">GC cycle</a> section
focused on the total CPU duration of an executing program.
However, such a duration is far less meaningful for say, a web service.
While throughput is still important for a web service (i.e. queries per second),
often the latency of each individual request matters even more.
</p>

<p>
In terms of latency, a stop-the-world GC may require a considerable length of
time to execute both its mark and sweep phases, during which the application,
and in the context of a web service, any in-flight request, is unable to make
further progress.
Instead, the Go GC avoids making the length of any global application pauses
proportional to the size of the heap, and that the core tracing algorithm is
performed while the application is actively executing.
(The pauses are more strongly proportional to GOMAXPROCS algorithmically, but
most commonly are dominated by the time it takes to stop running goroutines.)
Collecting concurrently is not without cost: in practice it often leads to a
design with lower throughput than an equivalent stop-the-world garbage
collector.
However,  it's important to note that <i>lower latency does not inherently mean
lower throughput</i>, and the performance of the Go garbage collector has
steadily improved over time, in both latency and throughput.
</p>

<p>
The concurrent nature of Go's current GC does not invalidate anything discussed
in this document so far: none of the statements relied on this design choice.
GC frequency is still the primary way the GC trades off between CPU
time and memory for throughput, and in fact, it also takes on this role for
latency.
This is because most of the costs for the GC are incurred while the mark phase
is active.
</p>

<p>
The key takeaway then, is that <b>reducing GC frequency may also lead to latency
improvements</b>.
This applies not only to reductions in GC frequency from modifying tuning
parameters, like increasing GOGC and/or the memory limit, but also applies to
the optimizations described in the
<a href="#Optimization_guide">optimization guide</a>.
</p>

<p>
However, latency is often more complex to understand than throughput, because it
is a product of the moment-to-moment execution of the program and not just an
aggregation of costs.
As a result, the connection between latency and GC frequency is less direct.
Below is a list of possible sources of latency for those inclined to dig
deeper.
</p>

<ol>
	<li>
		Brief stop-the-world pauses when the GC transitions between the mark
		and sweep phases,
	</li>
	<li>
		Scheduling delays because the GC takes 25% of CPU resources when in the
		mark phase,
	</li>
	<li>
		User goroutines assisting the GC in response to a high allocation rate,
	</li>
	<li>
		Pointer writes requiring additional work while the GC is in the mark
		phase, and
	</li>
	<li>
		Running goroutines must be suspended for their roots to be scanned.
	</li>
</ol>

<p>
These latency sources are visible in
<a href="/doc/diagnostics#execution-tracer">execution traces</a>, except for
pointer writes requiring additional work.
</p>

<!-- TODO: Add a short section about non-steady-state behavior. -->

<h3 id="Additional_resources">Additional resources</h3>

<p>
While the information presented above is accurate, it lacks the detail to
fully understand costs and trade-offs in the Go GC's design.
For more information, see the following additional resources.
</p>

<ul>
	<li>
		<a href="https://gchandbook.org/">The GC Handbook</a>&mdash;An
			excellent general resource and reference on garbage collector design.
	</li>
	<li>
		<a href="https://google.github.io/tcmalloc/design.html">TCMalloc</a>&mdash;Design
			document for the C/C++ memory allocator TCMalloc, which the Go memory allocator is based on.
	</li>
	<li>
		<a href="https://go.dev/blog/go15gc">Go 1.5 GC announcement</a>&mdash;The
			blog post announcing the Go 1.5 concurrent GC, which describes the algorithm in more detail.
	</li>
	<li>
		<a href="https://go.dev/blog/ismmkeynote">Getting to Go</a>&mdash;An
			in-depth presentation about the evolution of Go's GC design up to 2018.
	</li>
	<li>
		<a href="https://docs.google.com/document/d/1wmjrocXIWTr1JxU-3EQBI6BK6KgtiFArkG47XK73xIQ/edit">Go 1.5 concurrent GC pacing</a>&mdash;Design
			document for determining when to start a concurrent mark phase.
	</li>
	<li>
		<a href="https://github.com/golang/go/issues/30333">Smarter scavenging</a>&mdash;Design
			document for revising the way the Go runtime returns memory to the operating system.
	</li>
	<li>
		<a href="https://github.com/golang/go/issues/35112">Scalable page allocator</a>&mdash;Design
			document for revising the way the Go runtime manages memory it gets from the operating system.
	</li>
	<li>
		<a href="https://github.com/golang/go/issues/44167">GC pacer redesign (Go 1.18)</a>&mdash;Design
			document for revising the algorithm to determine when to start a concurrent mark phase.
	</li>
	<li>
		<a href="https://github.com/golang/go/issues/48409">Soft memory limit (Go 1.19)</a>&mdash;Design
			document for the soft memory limit.
	</li>
</ul>

<h2 id="A_note_about_virtual_memory">A note about virtual memory</h2>

<p>
This guide has largely focused on the physical memory use of the GC, but a
question that comes up regularly is what exactly that means and how it compares
to virtual memory (typically presented in programs like <code>top</code> as
"VSS").
</p>

<p>
Physical memory is memory housed in the actual physical RAM chip in most
computers.
<a href="https://en.wikipedia.org/wiki/Virtual_memory">Virtual memory</a> is an
abstraction over physical memory provided by the operating system to isolate
programs from one another.
It's also typically acceptable for programs to reserve virtual address space
that doesn't map to any physical addresses at all.
</p>

<p>
<b>
Because virtual memory is just a mapping maintained by the operating system,
it is typically very cheap to make large virtual memory reservations that don't
map to physical memory.
</b>
</p>

<p>
The Go runtime generally relies upon this view of the cost of virtual memory in
a few ways:
</p>

<ul>
	<li>
		<p>
		The Go runtime never deletes virtual memory that it maps.
		Instead, it uses special operations that most operating systems
		provide to explicitly release any physical memory resources
		associated with some virtual memory range.
		</p>

		<p>
		This technique is used explicitly to manage the
		<a href="#Memory_limit">memory limit</a> and return memory to the
		operating system that the Go runtime no longer needs.
		The Go runtime also releases memory it no longer needs continuously
		in the background.
		See <a href="#Additional_resources">the additional resources</a> for
		more information.
		</p>
	</li>
	<li>
		<p>
		On 32-bit platforms, the Go runtime reserves between 128 MiB and 512 MiB
		of address space up-front for the heap to limit fragmentation issues.
		</p>
	</li>
	<li>
		<p>
		The Go runtime uses large virtual memory address space reservations
		in the implementation of several internal data structures.
		On 64-bit platforms, these typically have a minimum virtual memory
		footprint of about 700 MiB.
		On 32-bit platforms, their footprint is negligible.
		</p>
	</li>
</ul>

<p>
As a result, virtual memory metrics such as "VSS" in <code>top</code> are
typically not very useful in understanding a Go program's memory footprint.
Instead, focus on "RSS" and similar measurements, which more directly reflect
physical memory usage.
</p>

<h2 id="Optimization_guide">Optimization guide</h2>

<h3 id="Identiying_costs">Identifying costs</h3>

<p>
Before trying to optimize how your Go application interacts with the GC, it's
important to first identify that the GC is a major cost in the first place.
</p>

</p>
The Go ecosystem provides a number of tools for identifying costs and optimizing
Go applications.
For a brief overview of these tools, see the
<a href="/doc/diagnostics">guide on diagnostics</a>.
Here, we'll focus on a subset of these tools and a reasonable order to apply
them in in order to understand GC impact and behavior.
</p>

<ol>
	<li>
		<p>
		<b>CPU profiles</b>
		</p>

		<p>
		A good place to start is with
		<a href="https://pkg.go.dev/runtime/pprof#hdr-Profiling_a_Go_program">CPU profiling</a>.
		CPU profiling provides an overview of where CPU time is spent, though to the
		untrained eye it may be difficult to identify the magnitude of the role the
		GC plays in a particular application.
		Luckily, understanding how the GC fits in mostly boils down to knowing what
		different functions in the `runtime` package mean.
		Below is a useful subset of these functions for interpreting CPU profiles.
		</p>

		<p>
		<i>
		Note: the functions listed below are not leaf functions, so they may not show
		up in the default the <code>pprof</code> tool provides with the <code>top</code>
		command.
		Instead, use the <code>top -cum</code> command or use the <code>list</code>
		command on these functions directly and focus on the cumulative percent column.
		</i>
		</p>
	</li>

	<ul>
		<li>
			<p>
			<b><code>runtime.gcBgMarkWorker</code></b>: Entrypoint to the
			dedicated mark worker goroutines.
			Time spent here scales with GC frequency and the complexity and
			size of the object graph.
			It represents a baseline for how much time the application spends
			marking and scanning.
			</p>
			<p>
			<i>
			Note: In a largely idle Go application, the Go GC is going to
			use up additional (idle) CPU resources to get its job done faster.
			As a result, this symbol may represent a large fraction of samples,
			that it believes are free.
			One common reason this can happen is if an application runs entirely
			in one goroutine but </i><code>GOMAXPROCS</code><i> is &gt;1.
			</i>
			</p>
		</li>
		<li>
			<p>
			<b><code>runtime.mallocgc</code></b>: Entrypoint to the memory
			allocator for heap memory.
			A large amount of cumulative time spent here (&gt;15%)
			typically indicates a lot of memory being allocated.
			</p>
		</li>
		<li>
			<p>
			<b><code>runtime.gcAssistAlloc</code></b>: Function goroutines
			enter to yield some of their time to assist the GC with scanning
			and marking.
			A large amount of cumulative time spent here (&gt;5%) indicates
			that the application is likely out-pacing the GC with respect to how
			fast it's allocating.
			It indicates a particularly high degree of impact from the GC,
			and also represents time the application spend marking and scanning.
			Note that this is included in the <code>runtime.mallocgc</code>
			call tree, so it will inflate that as well.
			</p>
		</li>
	</ul>

	<li>
		<p>
		<b>Execution traces</b>
		</p>

		<p>
		While CPU profiles are great for identifying where time is spent in
		aggregate, they're less useful for indicating performance costs that
		are more subtle, rare, or related to latency specifically.
		Execution traces</a> on
		the other hand provide a rich and deep view into a short window of
		a Go program's execution.
		They contain a variety of events related to the Go GC and specific
		execution paths can be directly observed, along with how the application
		might interact with the Go GC.
		All the GC events tracked are conveniently labeled as such in the trace
		viewer.
		</p>

		<p>
		See the <a href="https://pkg.go.dev/runtime/trace">documentation for the 
		<code>runtime/trace</code></a> package for how to get started with
		execution traces.
		</p>
	</li>

	<li>
		<p>
		<b>GC traces</b>
		</p>

		<p>
		When all else fails, the Go GC provides a few different specific traces
		that provide much deeper insights into GC behavior.
		These traces are always printed directly to STDERR, one line per GC cycle,
		and are configured through the <code>GODEBUG</code> environment variable
		that all Go programs recognize.
		They're mostly useful for debugging the Go GC itself since they require
		some familiarity with the specifics of the GC's implementation, but
		nonetheless can occasionally be useful to gain a better understanding of
		GC behavior.
		</p>

		<p>
		The core GC trace is enabled by setting <code>GODEBUG=gctrace=1</code>.
		The output produced by this trace is documented in the
		<a href="https://pkg.go.dev/runtime#hdr-Environment_Variables">environment
		variables section in the documentation for the <code>runtime</code>
		package.</a>
		</p>

		<p>
		A supplementary GC trace called the "pacer trace" provides even deeper
		insights and is enabled by setting <code>GODEBUG=gcpacertrace=1</code>.
		Interpreting this output requires an understanding of the GC's "pacer"
		(see <a href="#Additional_resources">additional resources</a>), which is
		outside the scope of this guide.
		</p>
	</li>
</ol>

<h3 id="Eliminating_heap_allocations">Eliminating heap allocations</h3>

<p>
One way to reduce costs from the GC is to have the GC manage fewer values to begin
with.
The techniques described below can produce some of the largest improvements in
performance, because as the <a href="#GOGC">GOGC section</a> demonstrated, the
allocation rate of a Go program is a major factor in GC frequency, the key
cost metric used by this guide.
</p>

<h4 id="Heap_profiling">Heap profiling</h4>

<p>
After <a href="#Identifying_costs">identifying that the GC is a source of
significant costs</a>, the next step in eliminating heap allocations is to
find out where most of them are coming from.
For this purpose, memory profiles (really, heap memory profiles) are very
useful.
Check out the <a href="https://pkg.go.dev/runtime/pprof#hdr-Profiling_a_Go_program">
documentation</a> for how to get started with them.
</p>

<p>
Memory profiles describe where in the program heap allocations come from,
identifying them by the stack trace at the point they were allocated.
Each memory profile can break down memory in four ways.
</p>

<ul>
	<li><code>inuse_objects</code>&mdash;Breaks down the number of objects that
	are live.</li>
	<li><code>inuse_space</code>&mdash;Breaks down live objects by how much
	memory they use in bytes.</li>
	<li><code>alloc_objects</code>&mdash;Breaks down the number of objects
	that have been allocated since the Go program began executing.</li>
	<li><code>alloc_space</code>&mdash;Breaks down the total amount of memory
	allocated since the Go program began executing.</li>
</ul>

<p>
Switching between these different views of heap memory may be done with either
the <code>-sample_index</code> flag to the <code>pprof</code> tool, or via the
<code>sample_index</code> option when the tool is used interactively.
</p>

<p>
<i>
Note: memory profiles by default only sample a subset of heap objects so they
will not contain information about every single heap allocation.
However, this is sufficient to find hot-spots.
To change the sampling rate, see
<a href="https://pkg.go.dev/runtime#pkg-variables"><code>runtime.MemProfileRate</code></a>.
</i>
</p>

<p>
For the purposes of reducing GC costs, <code>alloc_space</code> is typically the
most useful view as it directly corresponds to the allocation rate.
This view will indicate allocation hot spots that would provide the most benefit.
</p>

<h4 id="Escape_analysis">Escape analysis</h4>

<p>
Once candidate heap allocation sites have been identified with the help of
<a href="#Heap_profiling">heap profiles</a>, how can they be eliminated?
The key is to leverage the Go compiler's escape analysis to have the Go compiler
find alternative, and more efficient storage for this memory, for example
in the goroutine stack.
Luckily, the Go compiler has the ability to describe why it decides to escape
a Go value to the heap.
With that knowledge, it becomes a matter of reorganizing your source code to
change the outcome of the analysis (which is often the hardest part, but outside
the scope of this guide).
</p>

<p>
As for how to access the information from the Go compiler's escape analysis, the
simplest way is through a debug flag supported by the Go compiler that describes
all optimizations it applied or did not apply to some package in a text format.
This includes whether or not values escape.
Try the following command, where <code>[package]</code> is some Go package path.
</p>

<pre>
$ go build -gcflags=-m=3 [package]
</pre>

<p>
This information can also be visualized as an overlay in VS Code.
This overlay is configured and enabled in the VS Code Go plugin settings.
</p>

<ol>
	<li>
	Set the
	<a href="https://github.com/golang/vscode-go/wiki/settings#uicodelenses">
	<code>ui.codelenses</code> setting to include <code>gc_details</code>
	</a>.
	</li>
	<li>
	Enable the overlay for escape analysis by
	<a href="https://github.com/golang/vscode-go/wiki/settings#uidiagnosticannotations">
	setting <code>ui.diagnostic.annotations</code> to include <code>escape</code>
	</a>.
	</li>
</ol>

<p>
Finally, the Go compiler provides this information in a machine-readable (JSON)
format that may be used to build additional custom tooling.
For more information on that, see the
<a href="https://cs.opensource.google/go/go/+/master:src/cmd/compile/internal/logopt/log_opts.go;l=25;drc=351e0f4083779d8ac91c05afebded42a302a6893">
documentation in the source Go code</a>.
<p>

<h3 id="Implementation-specific_optimizations">Implementation-specific optimizations</h3>

<p>
The Go GC is sensitive to the demographics of live memory, because a complex
graph of objects and pointers both limits parallelism and generates more work
for the GC.
As a result, the GC contains a few optimizations for specific common structures.
The most directly useful ones for performance optimization are listed below.
</p>

<p>
<i>
Note: Applying the optimizations below may reduce the readability of your code
by obscuring intent, and may fail to hold up across Go releases.
Prefer to apply these optimizations only in the places they matter most.
Such places may be identified by using the tools listed in the
<a href="#Identifying_costs">section on identifying costs</a>.
</i>
</p>

<ul>
	<li>
		<p>
		Pointer-free values are segregated from other values.
		</p>
		<p>
		As a result, it may be advantageous to eliminate pointers from data
		structures that do not strictly need them, as this reduces the cache
		pressure the GC exerts on the program.
		As a result, data structures that rely on indices over pointer values,
		while less well-typed, may perform better.
		This is only worth doing if it's clear that the object graph is complex
		and the GC is spending a lot of time marking and scanning.
		</p>
	</li>
	<li>
		<p>
		The GC will stop scanning values at the last pointer in the value.
		</p>
		<p>
		As a result, it may be advantageous to group pointer fields in
		struct-typed values at the beginning of the value.
		This is only worth doing if it's clear the application spends a lot of its
		time marking and scanning.
		(In theory the compiler can do this automatically, but it is not yet
		implemented, and struct fields are arranged as written in the source
		code.)
		</p>
	</li>
</ul>

<p>
Furthermore, the GC must interact with nearly every pointer it sees, so using
indices into an slice, for example, instead of pointers, can aid in reducing GC
costs.
</p>

<h3 id="Linux_transparent_huge_pages">Linux transparent huge pages (THP)</h3>

<p>
When a program accesses memory, the CPU needs to translate the
<a href="#A_note_about_virtual_memory">virtual memory</a> addresses it uses
into physical memory addresses that refer to the data it was trying to access.
To do this, the CPU consults the "page table," a data structure that represents
the mapping from virtual to physical memory, managed by the operating system.
Each entry in the page table represents an indivisible block of physical memory
called a page, hence the name.
</p>

<p>
Transparent huge pages (THP) is a Linux feature that transparently replaces pages of
physical memory backing contiguous virtual memory regions with bigger blocks of memory
called huge pages.
By using bigger blocks, fewer page table entries are needed to represent the same memory
region, improving page table lookup times.
However, bigger blocks mean more waste if only a small part of the huge page is used
by the system.
</p>

<p>
When running Go programs in production, enabling transparent huge pages on Linux
can improve throughput and latency at the cost of additional memory use.
Applications with small heaps tend not to benefit from THP and may end up using a
substantial amount of additional memory (as high as 50%).
However, applications with big heaps (1 GiB or more) tend to benefit quite a bit
(up to 10% throughput) without very much additional memory overhead (1-2% or less).
Being aware of your THP settings in either case can be helpful, and experimentation
is always recommended.
</p>

<p>
One can enable or disable transparent huge pages in a Linux environment by modifying
<code>/sys/kernel/mm/transparent_hugepage/enabled</code>.
See the
<a href="https://www.kernel.org/doc/html/next/admin-guide/mm/transhuge.html">official
Linux admin guide</a> for more details.
If you choose to have your Linux production environment enable transparent huge pages,
we recommend the following additional settings for Go programs.
</p>

<ul>
	<li>
		<p>
		Set <code>/sys/kernel/mm/transparent_hugepage/defrag</code>
		to <code>defer</code> or <code>defer+madvise</code>.
		<br />
		<br />
		This setting controls how aggressively a Linux kernel coalesces regular
		pages into huge pages.
		<code>defer</code> tells the kernel to coalesce huge pages lazily
		and in the background.
		A more aggressive setting can induce stalls in memory constrained systems
		and can often hurt application latencies.
		<code>defer+madvise</code> is like <code>defer</code>, but is friendlier
		to other applications on the system that request huge pages explicitly and
		require them for performance.
		</p>
	</li>
	<li>
		<p>
		Set <code>/sys/kernel/mm/transparent_hugepage/khugepaged/max_ptes_none</code>
		to <code>0</code>.
		<br />
		<br />
		This setting controls how many additional pages the Linux kernel daemon
		can allocate when trying to allocate a huge page.
		The default setting is maximally aggressive, and can often
		<a href="https://bugzilla.kernel.org/show_bug.cgi?id=93111">undo work the Go
		runtime does to return memory to the OS</a>.
		Before Go 1.21, the Go runtime tried to mitigate the negative effects of the
		default setting, but it came with a CPU cost.
		With Go 1.21+ and Linux 6.2+, the Go runtime will coalesce regular pages
		into huge pages itself with its own more accurate heuristics.
		</p>
	</li>
</ul>

<h2 id="Appendix">Appendix</h2>

<h3 id="Additional_notes_on_GOGC">Additional notes on GOGC</h3>

<p>
The <a href="#GOGC">GOGC section</a> claimed that doubling GOGC doubles heap
memory overheads and halves GC CPU costs.
To see why, let's break it down mathematically.
</p>

<p>
Firstly, the heap target sets a target for the total heap size.
This target, however, mainly influences the new heap memory, because the live
heap is fundamental to the application.
</p>

<p class="gc-guide-equation">
<i>
Target heap memory = Live heap + (Live heap + GC roots) * GOGC / 100
</i>
</p>

<p class="gc-guide-equation">
<i>
Total heap memory = Live heap + New heap memory
</i>
</p>

<p class="gc-guide-equation">
<i>
&rArr;
</i>
</p>

<p class="gc-guide-equation">
<i>
New heap memory = (Live heap + GC roots) * GOGC / 100
</i>
</p>

<p>
From this we can see that doubling GOGC would also double the amount of new heap
memory that application will allocate each cycle, which captures heap memory
overheads.
Note that <i>Live heap + GC roots</i> is an approximation of the amount of
memory the GC needs to scan.
</p>

<p>
Next, let's look at GC CPU cost.
Total cost can be broken down as the cost per cycle, times GC frequency over
some time period T.
</p>

<p class="gc-guide-equation">
<i>
Total GC CPU cost = (GC CPU cost per cycle) * (GC frequency) * T
</i>
</p>

<p>
GC CPU cost per cycle can be derived from the
<a href="#Understanding_costs">GC model</a>:
</p>

<p class="gc-guide-equation">
<i>
GC CPU cost per cycle = (Live heap + GC roots) * (Cost per byte) + Fixed cost
</i>
</p>

<p>
Note that sweep phase costs are ignored here as mark and scan costs dominate.
</p>

<p>
The steady-state is defined by a constant allocation rate and a constant cost
per byte, so in the steady-state we can derive a GC frequency from this new heap
memory:
</p>

<p class="gc-guide-equation">
<i>
GC frequency = (Allocation rate) / (New heap memory) = (Allocation rate) / ((Live heap + GC roots) * GOGC / 100)
</i>
</p>

<p>
Putting this together, we get the full equation for the total cost:
</p>

<p class="gc-guide-equation">
<i>
Total GC CPU cost = (Allocation rate) / ((Live heap + GC roots) * GOGC / 100) * ((Live heap + GC roots) * (Cost per byte) + Fixed cost) * T
</i>
</p>

<p>
For a sufficiently large heap (which represents most cases), the marginal costs
of a GC cycle dominate the fixed costs.
This allows for a significant simplification of the total GC CPU cost formula.
</p>

<p class="gc-guide-equation">
<i>
Total GC CPU cost = (Allocation rate) / (GOGC / 100) * (Cost per byte) * T
</i>
</p>

<p>
From this simplified formula, we can see that if we double GOGC, we halve total
GC CPU cost.
(Note that the visualizations in this guide do simulate fixed costs, so the GC
CPU overheads reported by them will not exactly halve when GOGC doubles.)
Furthermore, GC CPU costs are largely determined by allocation rate and the
cost per byte to scan memory.
For more information on how to reduce these costs specifically, see the
<a href="#Optimization_guide">optimization guide</a>.
</p>

<p>
<i>
Note: there exists a discrepancy between the size of the live heap, and the
amount of that memory the GC actually needs to scan: the same size live heap but
with a different structure will result in a different CPU cost, but the same
memory cost, resulting a different trade-off.
This is why the structure of the heap is part of the definition of the
steady-state.
The heap target should arguably only include the scannable live heap as a closer
approximation of memory the GC needs to scan, but this leads to degenerate
behavior when there's a very small amount of scannable live heap but the live
heap is otherwise large.
</i>
</p>

<script src="/js/d3.js"></script>
<script async src="/doc/gc-guide.js"></script>