A Guide to the Go Garbage Collector

Introduction

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.

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 physical memory 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.

Another term for automatically recycling memory is garbage collection. At a high level, a garbage collector (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.

Note that the existence of a garbage collector as described by this guide is not guaranteed by the Go specification, 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.

Therefore, this guide is about a specific implementation of the Go programming language and may not apply to other implementations. Specifically, this following guide applies to the standard toolchain (the gc 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.

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.

Where Go Values Live

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

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 lexical scope 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.

Go values whose memory cannot be allocated this way, because the Go compiler cannot determine its lifetime, are said to escape to the heap. "The heap" can be thought of as a catch-all for memory allocation, for when Go values need to be placed somewhere. 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.

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.

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 eliminating heap allocations.

Tracing Garbage Collection

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 tracing garbage collection, which identifies in-use, so-called live, objects by following pointers transitively.

Let's define these terms more rigorously.

• Object—An object is a dynamically allocated piece of memory that contains one or more Go values.

• Pointer—A memory address that references any value within an object. This naturally includes Go values of the form *T, 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.

Together, objects and pointers to other objects form the object graph. To identify live memory, the GC walks the object graph starting at the program's roots, 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 scanning.

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 is that in order to keep track of its progress, the GC also marks 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 not marked available for allocation. This process is called sweeping.

One alternative technique you may be familiar with is to actually move 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 moving GC; Go has a non-moving GC.

The GC cycle

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 all 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 continuosly rotates through these three states of off, marking, and sweeping in what's known as the GC cycle.

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.

Understanding costs

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.

To begin with, consider this model of GC cost based on four simple axioms.

1. The application is paused while the GC executes.

2. The GC involves only two resources: CPU time, and physical memory.

3. 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.

   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.

4. 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.

   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.

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 steady-state.

• The rate at which the application allocates new memory (in bytes per second) is constant.

  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.)

  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. 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.

• 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).

  Another way to think about this is that the marginal costs of GC are constant.

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.

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 GOGC section explores the non-constant live heap scenario in some more detail.

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.

Now consider if the GC were to shift the point at which it runs later 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 earlier in time: less memory would be allocated and CPU costs would be incurred more often.

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 GC frequency.

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.

GOGC is a tuning parameter for the Go GC that directly reflects a trade-off between CPU time and memory by controlling GC frequency. More specifically, GOGC sets the target heap size for the GC, or the amount of new memory that should be allocated by the time the mark phase has completed. GOGC is defined as a percent overhead over the amount of work the GC needs to to do. This work is currently defined as the size of the live heap, plus the size of the GC roots in bytes.

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.

Note: GOGC may be more precisely described as defining the amount of new memory that can be allocated before the start of the next sweep phase. This timing is technically true for the GC model this guide has been using thus far, but also generalizes to the real GC implementation that Go uses that's discussed in more detail in the latency section.

The benefit of defining the trade-off this way is that the cost of GC remains constant in the steady-state regardless of the amount of work that the GC has to do (so regardless of the sizes of the live heap and root set), because the frequency will always be proportional to the amount of work that has to be done. In other words, it represents a fixed point in the trade-off between CPU cost and memory use. (It's important to note that this fixed point may shift if the steady-state changes as well, but crucially it is not dependent on the size of the live heap.)

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.

GOGC may be configured through either the GOGC environment variable (which all Go programs recognize), or through the SetGCPercent API in the runtime/debug package.

Note that GOGC may also be used to turn off the GC entirely (provided the memory limit does not apply) by setting GOGC=off or calling SetGCPercent(-1). 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.

To better understand everything we've discussed so far, try out the interactive visualization below that is built on the GC cost model 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.

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 occurs while the new heap drops to zero. 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.

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.

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.

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.

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.

Memory limit

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 peak live heap size, even if in the usual case a higher GOGC value provides a better trade-off.

The visualization below demonstrates this transient heap spike situation.

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.

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 GOMEMLIMIT environment variable which all Go programs recognize, or through the SetMemoryLimit function available in the runtime/debug package.

This memory limit sets a maximum on the total amount of memory that the Go runtime can use. The specific set of memory included is defined in terms of runtime.MemStats as the expression

Sys - HeapReleased

or equivanently in terms of the runtime/metrics package,

/memory/classes/total:bytes - /memory/classes/heap/released:bytes

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.

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.

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.

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.

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.

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 maximization of resource economy because it sets the minimum GC frequency required to maintain some memory limit. In this case, all of the program's execution has the heap size rise to meet the memory limit.

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

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.

This situation, where the program fails to make reasonable progress due to constant GC cycles, is called thrashing. 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.

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

For this reason, the memory limit is defined to be soft. 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.

How this works internally is the GC mitigates 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 2 * GOMAXPROCS 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.

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.

Note: the visualizations on this page do not simulate the GC CPU limit.

Suggested uses

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.

• Do 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).

  A good example is the deployment of a web service into containers with a fixed amount of available memory.

  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.

• Do feel free to adjust the memory limit in real time to adapt to changing conditions.

  A good example is a cgo program where C libraries temporarily need to use substantially more memory.

• Don't 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.

  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.

• Don't 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.

  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.

• Don't set a memory limit to avoid out-of-memory conditions when a program is already close to its environment's memory limits.

  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 then potentially set a memory limit) or decrease GOGC (which provides a much cleaner trade-off than thrashing-mitigation does).

Latency

Until this point, this document has 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.

The Go GC, however, is not fully stop-the-world, and in fact does most of its work concurrently with the application. The main reason for this is that it reduces application latencies. Specifically, the end-to-end duration of a single unit of computation (e.g. a web request). Thus far, this document mainly considered application throughput, or the aggregation of these operations (e.g. web requests handled per second). Note that each example in the GC cycle section focused on the total CPU duration of an executing program. However, such a duration is far less meaningful for say, a web service, whose duration primarily captures reliability (i.e. uptime) and not cost. While throughput is still important for a web service (i.e. queries per second), often the latency of each individual request matters even more, as it correlates with other important metrics.

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 ensures that the length of any global application pauses are never proportional to the size of the heap in any form, and that the core tracing algorithm is performed while the application is actively executing. This choice is not without cost, as in practice it tends to lead to a design with lower throughput, but it's important to note that low latency does not inherently mean low throughput, even though in many cases the two are at odds with one another.

At first, the concurrent nature of the Go GC may appear to be a significant departure from the cost model presented earlier. Fortunately, the intuition behind the model still applies.

Although the first axiom no longer holds, it wasn't really all that important to begin with; the rest of the costs still align as described by the model, and the same notion of a steady-state applies. As a result, GC frequency is still the primary way the GC trades off between CPU time and memory for throughput, and it also takes on this role for latency. With respect to throughput, it's easy to get back within the realm of the model by just pretending all the little costs the concurrent GC incurs happened at the end of the GC cycle. With respect to latency, most of the added latency from the GC comes specifically from the period of time when the mark phase is active. Thus, the more often the GC is in the mark phase, the more often these costs are incurred, and so latency also follows GC frequency.

More concretely, adjusting GC tuning parameters to reduce GC frequency may also lead to latency improvements. That means increasing GOGC and/or the memory limit.

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 more tenuous and may not be quite as direct. Below is a list of possible sources of latency for those inclined to dig deeper. These latency sources are visible in execution traces.

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

Additional resources

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.

• The GC Handbook—An excellent general resource and reference on garbage collector design.
• TCMalloc—Design document for the C/C++ memory allocator TCMalloc, which the Go memory allocator is based on.
• Go 1.5 GC announcement—The blog post announcing the Go 1.5 concurrent GC, which describes the algorithm in more detail.
• Getting to Go—An in-depth presentation about the evolution of Go's GC design up to 2018.
• Go 1.5 concurrent GC pacing—Design document for determining when to start a concurrent mark phase.
• Smarter scavenging—Design document for revising the way the Go runtime returns memory to the operating system.
• Scalable page allocator—Design document for revising the way the Go runtime manages memory it gets from the operating system.
• GC pacer redesign (Go 1.18)—Design document for revising the algorithm to determine when to start a concurrent mark phase.
• Soft memory limit (Go 1.19)—Design document for the soft memory limit.

A note about virtual memory

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 top as "VSS").

Physical memory is memory housed in the actual physical RAM chip in most computers. Virtual memory 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.

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.

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

• 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.

  This technique is used explicitly to manage the memory limit 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 the additional resources for more information.

• 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.

• 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.

As a result, virtual memory metrics such as "VSS" in top 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.

Optimization guide

Identifying costs

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.

The Go ecosystem provides a number of tools for identifying costs and optimizing Go applications. For a brief overview of these tools, see the guide on diagnostics. 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.

1. CPU profiles

   A good place to start is with CPU profiling. 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.

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

• runtime.gcBgMarkWorker: 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.

  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 GOMAXPROCS is >1.

• runtime.mallocgc: Entrypoint to the memory allocator for heap memory. A large amount of cumulative time spent here (>15%) typically indicates a lot of memory being allocated.

• runtime.gcAssistAlloc: 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 (>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 runtime.mallocgc call tree, so it will inflate that as well.

Execution traces

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 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.

See the documentation for the runtime/trace package for how to get started with execution traces.

GC traces

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 GODEBUG 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.

The core GC trace is enabled by setting GODEBUG=gctrace=1. The output produced by this trace is documented in the environment variables section in the documentation for the runtime package.

A supplementary GC trace called the "pacer trace" provides even deeper insights and is enabled by setting GODEBUG=gcpacertrace=1. Interpreting this output requires an understanding of the GC's "pacer" (see additional resources), which is outside the scope of this guide.

Eliminating heap allocations

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 GOGC section demonstrated, the allocation rate of a Go program is a major factor in GC frequency, the key cost metric used by this guide.

Heap profiling

After identifying that the GC is a source of significant costs, 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 documentation for how to get started with them.

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.

• inuse_objects—Breaks down the number of objects that are live.
• inuse_space—Breaks down live objects by how much memory they use in bytes.
• alloc_objects—Breaks down the number of objects that have been allocated since the Go program began executing.
• alloc_space—Breaks down the total amount of memory allocated since the Go program began executing.

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

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 runtime.MemProfileRate.

For the purposes of reducing GC costs, alloc_space 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.

Escape analysis

Once candidate heap allocation sites have been identified with the help of heap profiles, 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).

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 [package] is some Go package path.

$ go build -gcflags=-m=3 [package]

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.

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 documentation in the source Go code.

Implementation-specific optimizations

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.

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 section on identifying costs.

• Pointer-free values are segregated from other values.

  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.

• The GC will stop scanning values at the last pointer in the value.

  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.)

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.