Go Memory Management

This is a blog post version of a talk I gave at Vilnius Go Meetup. If you’re ever in Vilnius and enjoy Go come join us and consider speaking ūüôā

So, in this post we will explore Go memory management. Let’s begin with a following little program:

func main() {
http.HandleFunc("/bar", func(w http.ResponseWriter, r *http.Request) {
fmt.Fprintf(w, "Hello, %q", html.EscapeString(r.URL.Path))

http.ListenAndServe(":8080", nil)

Let’s compile and run it:

go build main.go

Now let’s find the running process via ps:

ps -u --pid 16609
povilasv 16609 0.0 0.0 388496 5236 pts/9 Sl+ 17:21 0:00 ./main

We can see that this program consumes 379.39 MiB of virtual memory and resident size is 5.11 mb. Wait what? Why ~380 MiB?

A little reminder:

Virtual Memory Size(VSZ) is all memory that the process can access, including memory that is swapped out, memory that is allocated, but not used, and memory that is from shared libraries. (Edited, good explanation in stackoverflow.)

Resident Set Size(RSS) is number of memory pages the process has in real memory multiplied by pagesize. This excludes swapped out memory pages.

Before deep diving into this problem, let’s go thru some basics of computer architecture and memory management in computers.

Memory Basics

Wikipedia defines RAM as:

Random-access memory (RAM /ræm/) is a form of computer data storage that stores data and machine code currently being used.
A random-access memory device allows data items to be read or written in almost the same amount of time irrespective of the physical location of data inside the memory.

I view physical memory as this array of slots/units, where slot can hold 8 bits of information 1. Each memory slot has an address and in your programs you tell CPU “yo CPU, could you grab that byte of information out of memory at address 0“, or “yo CPU, could you put this byte of information at address 1“.

As computers typically are running multiple tasks, reading and writing directly from/to physical memory is a bad idea. Imagine how easy is to write a program, which reads all the stuff (including your passwords) out of memory or a program, which would write at different program’s memory addresses. That would be wild wild west.

So, instead of doing things with Physical Memory we have a concept of Virtual Memory. When your program runs, it only sees it’s memory and it thinks that it’s the only one in here 2. Also, not all of your program’s stored memory bytes could be in RAM. If you don’t access specific memory block often enough, Operating System can put some block of memory into slower storage (like disk) saving precious RAM. And OS won’t even admit to your application that OS did it. But we all know that OS did it.

Virtual memory can be implemented using Segmentation or Page tables based on your CPU architecture and OS. I’m not going to go into detail about Segmentation as Page tables are way more common, but you can read more about Segmentation in 3.

In Paged Virtual Memory, we divide virtual memory into blocks, called Pages. Pages can vary in size based on hardware, but usually pages are 4-64 KB, often with the capability to use huge pages from 2 MB to 1 GB. The division into blocks is useful as it would require a lot more memory to manage each memory slot individually and would slow down performance of your computer.

In order to implement Paged Virtual Memory, there is a chip called Memory Magament Unit (MMU) 4, which sits between CPU and your memory. MMU holds mapping from virtual address to physical address in a table (which it stores in memory) called Page Table, containing one Page Table Entry (PTE) per page. Also MMU has a physical cache called Translation Lookaside Buffer (TLB), which store recent translations from Virtual Memory to Physical. Schematically it looks like this:

So let’s say OS decides to put some virtual memory page into disk and your program tries to access it. This process looks like this:

  1. CPU issues a command to access the virtual address, MMU check’s it in it’s Page Table and prohibits access, because no Physical RAM has been allocated to that virtual page.
  2. Then MMU sends a Page Fault to the CPU.
  3. The Operating System then handles the Page fault, by finding a spare memory block of RAM (called frame) and setting up new PTE to map it.
  4. If no RAM is free, it may choose an existing page, using some replacement algorithm, and save it to disk (this process is called paging).
  5. With some Memory Management Units, there can also be a shortage of Page Table Entrys, in which case the OS will have to free one for the new mapping.

Operating systems usually manages multiple applications(processes) so the whole memory management bit looks like this:

Each process has one linear virtual address space, with addresses running from 0 to some huge maximum. Virtual address space doesn’t need to be contiguous, so that not all of these virtual addresses are actually used to store data and they don’t consume space in RAM or disk. What’s really cool about this is that the same frame of real memory can back multiple virtual pages belonging to multiple processes. This is normally the case, with virtual memory occupied by GNU C Library code (libc), which is included by default if you comple with go build. You can compile go code without libc via ldflags setting 5:

go build -ldflags '-libgcc=none'

So this is a high level overview of what memory is and how it’s implemented using hardware and OS talking to each other. Now let’s take a look at what happens in OS, when you try to run your program and how programs can allocate memory.

Operating system

In order to run programs Operating System has a module, which loads programs and needed libraries, called program loader. In Linux you can call program loader from your program using execve() system call.

When loader runs it goes thru these steps 6:

  1. validates the program image (permissions, memory requirements etc.);
  2. copies the program image from the disk into main memory;
  3. passes the command-line arguments on the stack;
  4. initializes registers (like the stack pointer);

Once loading is complete, the operating system starts the program by passing control to the loaded program code (executing a jump instruction to program’s entrypoint (_start)).

So what is a program anyway?

Usually you write programs in high-level languages like Go, which are compiled into either an executable machine code file or a non-executable machine-code object file (a library). These Object files, executable or not, are typically in a container format, such as Executable and Linkable Format (ELF) (typically in Linux), Portable Executable (typically in Windows). Sometimes you don’t have a luxury of writing everything in beloved Go. In that case, one option is to hand craft your own ELF binary and put machine code into correct ELF structure. Another option is to develop a program in assembly language, which remains human-readable while being more closely associated with machine code instructions.

Object files are binary representations of program intended to execute directly on to processor. Those object files contain not only machine code, but also carry metadata about application, like OS architecture, debug information. Also object files carry application data, like global variables or constants. Typically object files are structured into sections such as the .text (executable code), .data (global variables), and .rodata (global constants) 7.

So I’m running Linux (Ubuntu) and my compiled programs (output file of go build) are in ELF 8. In Go, we can easily write a program, which reads an ELF executable file, because Go has a debug/elf package in standard library. Here is an example:

package main

import (

func main() {
f, err := elf.Open("main")

if err != nil {

for _, section := range f.Sections {

And the output:

2018/05/06 14:26:08 &{{ SHT_NULL 0x0 0 0 0 0 0 0 0 0} 0xc4200803f0 0xc4200803f0 0 0}
2018/05/06 14:26:08 &{{.text SHT_PROGBITS SHF_ALLOC+SHF_EXECINSTR 4198400 4096 3373637 0 0 16 0 3373637} 0xc420080420 0xc420080420 0 0}
2018/05/06 14:26:08 &{{.plt SHT_PROGBITS SHF_ALLOC+SHF_EXECINSTR 7572064 3377760 560 0 0 16 16 560} 0xc420080450 0xc420080450 0 0}
2018/05/06 14:26:08 &{{.rodata SHT_PROGBITS SHF_ALLOC 7573504 3379200 1227675 0 0 32 0 1227675} 0xc420080480 0xc420080480 0 0}
2018/05/06 14:26:08 &{{.rela SHT_RELA SHF_ALLOC 8801184 4606880 24 11 0 8 24 24} 0xc4200804b0 0xc4200804b0 0 0}
2018/05/06 14:26:08 &{{.rela.plt SHT_RELA SHF_ALLOC 8801208 4606904 816 11 2 8 24 816} 0xc4200804e0 0xc4200804e0 0 0}
2018/05/06 14:26:08 &{{.gnu.version SHT_GNU_VERSYM SHF_ALLOC 8802048 4607744 78 11 0 2 2 78} 0xc420080510 0xc420080510 0 0}
2018/05/06 14:26:08 &{{.gnu.version_r SHT_GNU_VERNEED SHF_ALLOC 8802144 4607840 112 10 2 8 0 112} 0xc420080540 0xc420080540 0 0}
2018/05/06 14:26:08 &{{.hash SHT_HASH SHF_ALLOC 8802272 4607968 192 11 0 8 4 192} 0xc420080570 0xc420080570 0 0}
2018/05/06 14:26:08 &{{.shstrtab SHT_STRTAB 0x0 0 4608160 375 0 0 1 0 375} 0xc4200805a0 0xc4200805a0 0 0}
2018/05/06 14:26:08 &{{.dynstr SHT_STRTAB SHF_ALLOC 8802848 4608544 594 0 0 1 0 594} 0xc4200805d0 0xc4200805d0 0 0}
2018/05/06 14:26:08 &{{.dynsym SHT_DYNSYM SHF_ALLOC 8803456 4609152 936 10 0 8 24 936} 0xc420080600 0xc420080600 0 0}
2018/05/06 14:26:08 &{{.typelink SHT_PROGBITS SHF_ALLOC 8804416 4610112 12904 0 0 32 0 12904} 0xc420080630 0xc420080630 0 0}
2018/05/06 14:26:08 &{{.itablink SHT_PROGBITS SHF_ALLOC 8817320 4623016 3176 0 0 8 0 3176} 0xc420080660 0xc420080660 0 0}
2018/05/06 14:26:08 &{{.gosymtab SHT_PROGBITS SHF_ALLOC 8820496 4626192 0 0 0 1 0 0} 0xc420080690 0xc420080690 0 0}
2018/05/06 14:26:08 &{{.gopclntab SHT_PROGBITS SHF_ALLOC 8820512 4626208 1694491 0 0 32 0 1694491} 0xc4200806c0 0xc4200806c0 0 0}
2018/05/06 14:26:08 &{{.got.plt SHT_PROGBITS SHF_WRITE+SHF_ALLOC 10518528 6324224 296 0 0 8 8 296} 0xc4200806f0 0xc4200806f0 0 0}
2018/05/06 14:26:08 &{{.dynamic SHT_DYNAMIC SHF_WRITE+SHF_ALLOC 10518848 6324544 304 10 0 8 16 304} 0xc420080720 0xc420080720 0 0}
2018/05/06 14:26:08 &{{.got SHT_PROGBITS SHF_WRITE+SHF_ALLOC 10519152 6324848 8 0 0 8 8 8} 0xc420080750 0xc420080750 0 0}
2018/05/06 14:26:08 &{{.noptrdata SHT_PROGBITS SHF_WRITE+SHF_ALLOC 10519168 6324864 183489 0 0 32 0 183489} 0xc420080780 0xc420080780 0 0}
2018/05/06 14:26:08 &{{.data SHT_PROGBITS SHF_WRITE+SHF_ALLOC 10702688 6508384 46736 0 0 32 0 46736} 0xc4200807b0 0xc4200807b0 0 0}
2018/05/06 14:26:08 &{{.bss SHT_NOBITS SHF_WRITE+SHF_ALLOC 10749440 6555136 127016 0 0 32 0 127016} 0xc4200807e0 0xc4200807e0 0 0}
2018/05/06 14:26:08 &{{.noptrbss SHT_NOBITS SHF_WRITE+SHF_ALLOC 10876480 6682176 12984 0 0 32 0 12984} 0xc420080810 0xc420080810 0 0}
2018/05/06 14:26:08 &{{.tbss SHT_NOBITS SHF_WRITE+SHF_ALLOC+SHF_TLS 0 0 8 0 0 8 0 8} 0xc420080840 0xc420080840 0 0}
2018/05/06 14:26:08 &{{.debug_abbrev SHT_PROGBITS 0x0 10891264 6557696 437 0 0 1 0 437} 0xc420080870 0xc420080870 0 0}
2018/05/06 14:26:08 &{{.debug_line SHT_PROGBITS 0x0 10891701 6558133 350698 0 0 1 0 350698} 0xc4200808a0 0xc4200808a0 0 0}
2018/05/06 14:26:08 &{{.debug_frame SHT_PROGBITS 0x0 11242399 6908831 381068 0 0 1 0 381068} 0xc4200808d0 0xc4200808d0 0 0}
2018/05/06 14:26:08 &{{.debug_pubnames SHT_PROGBITS 0x0 11623467 7289899 121435 0 0 1 0 121435} 0xc420080900 0xc420080900 0 0}
2018/05/06 14:26:08 &{{.debug_pubtypes SHT_PROGBITS 0x0 11744902 7411334 225106 0 0 1 0 225106} 0xc420080930 0xc420080930 0 0}
2018/05/06 14:26:08 &{{.debug_gdb_scripts SHT_PROGBITS 0x0 11970008 7636440 53 0 0 1 0 53} 0xc420080960 0xc420080960 0 0}
2018/05/06 14:26:08 &{{.debug_info SHT_PROGBITS 0x0 11970061 7636493 1847750 0 0 1 0 1847750} 0xc420080990 0xc420080990 0 0}
2018/05/06 14:26:08 &{{.debug_ranges SHT_PROGBITS 0x0 13817811 9484243 167568 0 0 1 0 167568} 0xc4200809c0 0xc4200809c0 0 0}
2018/05/06 14:26:08 &{{.interp SHT_PROGBITS SHF_ALLOC 4198372 4068 28 0 0 1 0 28} 0xc4200809f0 0xc4200809f0 0 0}
2018/05/06 14:26:08 &{{.note.go.buildid SHT_NOTE SHF_ALLOC 4198272 3968 100 0 0 4 0 100} 0xc420080a20 0xc420080a20 0 0}
2018/05/06 14:26:08 &{{.symtab SHT_SYMTAB 0x0 0 9654272 290112 35 377 8 24 290112} 0xc420080a50 0xc420080a50 0 0}
2018/05/06 14:26:08 &{{.strtab SHT_STRTAB 0x0 0 9944384 446735 0 0 1 0 446735} 0xc420080a80 0xc420080a80 0 0}

You can also inspect ELF files using Linux tools like: size --format=sysv main or readelf -l main (here main is the output binary).

As you can see that executable is just a file with some predifined format. Usually executable formats have segments, which are blocks of data memory mapped before running image. Here is a common view of segements, that a process has:

The text segment contains a program’s instructions and literals and static constants.

The data segment is working storage for the program. It can be preallocated and preloaded by exec and the process can extend or shrink it.

The stack segment contains a program stack. It grows as the stack grows, but doesn’t shrink when the stack shrinks.

The heap area commonly begins at the end of the .bss and .data segments and grows to larger addresses from there.

Let’s take look at how processes can allocate memory.

Libc manual says 9, that programs can allocate in two major ways using exec family functions and programmatically. exec calls out program loader to launch a program, thus creating a virtual address space for a process, loading its program into it and running it. And the programmatic ways are:

  • Static allocation is what happens when you declare a global variable. Each global variable defines one block of space, of a fixed size. The space is allocated once, when your program is started (part of the exec operation), and is never freed.

  • Automatic allocation happens when you declare an automatic variable, such as a function argument or a local variable. The space for an automatic variable is allocated when the compound statement containing the declaration is entered, and is freed when that compound statement is exited.

  • Dynamic Allocation – is a technique in which programs determine as they are running where to store some information. You need dynamic allocation when the amount of memory you need, or how long you continue to need it, depends on factors that are not known before the program runs.

To dynamically allocate memory you have a couple of options. One of the options is to make a call to operating system (syscall or via libc). OS provides various functions like:

  • mmap/munmap – allocates/deallocates fixed block memory page.
  • brk/sbrk – changes/gets data segement size
  • madvise – gives advise to Operating System how to manage memory
  • set_thread_area/get_thread_area – works with thread local storage.

I think Go runtime only uses mmap, madvise, munmap and sbrk and it calls it directly to underlying OS via assembly or cgo, i.e. it’s not calling out to libc 10. These memory allocation are low level and typically programers don’t use them. More common is to use libc’s malloc family functions, where you ask for n bytes of memory and libc just returns it to you, and you need to call free to return it back.

Here is a basic C example of using malloc:

#include /* printf, scanf, NULL */
#include /* malloc, free, rand */
int main (){
int i,n;
char * buffer;

printf ("How long do you want the string? ");
scanf ("%d", &i);

buffer = (char*) malloc (i+1);
if (buffer==NULL) exit (1);

for (n=0; n<i; n++)

printf ("Random string: %s\n",buffer);
free (buffer);

return 0;

This example illustrates the need to dynamically allocate data, as we ask the user to enter string length, then allocate bytes based on that and generate random string. Also, note the explicit call to free()

Memory Allocators

As Go doesn’t use malloc to get memory, but asks OS directly (via mmap), it has to implement memory allocation and deallocation on it’s own (like malloc does). Go’s memory allocator is originally based off TCMalloc: Thread-Caching Malloc.

Some interesting facts about TCMalloc:

  • TCMalloc is faster than the glibc 2.3 malloc (available as a separate library called ptmalloc2).
  • ptmalloc2 takes approximately 300 nanoseconds to execute a malloc.
  • The TCMalloc implementation takes approximately 50 nanoseconds for the same operation pair.

TCMalloc also reduces lock contention for multi-threaded programs:

  • For small objects, there is virtually zero contention.
  • For large objects, TCMalloc tries to use fine grained and efficient spinlocks.


The secret behind TCMalloc performance is that it uses thread-local cache to store some preallocated memory “objects”, so that small allocations are satisfied from the thread-local cache 11. Once thread-local cache is out of space, memory objects are moved from central data structures into thread-local cache.

TCMalloc treats small object (size <= 32K) allocation differently from large. Large objects are allocated directly from the central heap using a page-level allocator. While, small objects are maped to one of approximately 170 allocatable size-classes.

So here is how it works for small objects:

When allocating a small object:

  1. We map its size to the corresponding size-class.
  2. Look in the corresponding free list in the thread cache for the current thread.
  3. If the free list is not empty, we remove the first object from the list and return it.

If the free list is empty:

  1. We fetch a bunch of objects from a central free list for this size-class (the central free list is shared by all threads).
  2. Place them in the thread-local free list.
  3. Return one of the newly fetched objects to the applications.

If the central free list is also empty:

  1. We allocate a run of pages from the central page allocator.
  2. Split the run into a set of objects of this size-class.
  3. Place the new objects on the central free list.
  4. As before, move some of these objects to the thread-local free list.

Large Objects (size > 32K) is rounded up to a page size (4K) and is handled by a central page heap. The central page heap is again an array of free lists:

For i < 256, the kth entry is a free list of runs that consist of k pages. The 256th entry is a free list of runs that have length >= 256 pages.

So here is how it works for large objects:

An allocation for k pages is satisfied:

  1. We look in the k-th free list.
  2. If that free list is empty, we look in the next free list, and so forth.
  3. Eventually, we look in the last free list if necessary.
  4. If that fails, we fetch memory from the system.
  5. If an allocation for k pages is satisfied by a run of pages of length > k, the remainder of the run is re-inserted back into the appropriate free list in the page heap.

The memory is managed in term of runs of contiguous pages, which are called Spans (this is important as Go also manages memory in terms of Spans).

In TCMalloc a span can either be allocated, or free:

  • If free, the span is one of the entries in a page heap linked-list.
  • If allocated, it is either a large object that has been handed off to the application, or a run of pages that have been split up into a sequence of small objects.

In this example, span 1 occupies 2 pages, span 2 occupies 4 pages, span 3 occupies 1 page. A central array indexed by page number can be used to find the span to which a page belongs.

Go Memory Allocator

Go allocator is similiar to TCMalloc, it works in runs of pages (spans / mspan objects), uses thread-local cache and divides allocations based on size. Spans are contiguous regions of memory of 8K or larger. You can see Span for yourself in in runtime/mheap.go there is mspan struct. There are 3 types of Spans:

  1. idle ‚Äď span, that has no objects and can be released back to the OS, or reused for heap allocation, or reused for stack memory.
  2. in use ‚Äď span, that has atleast one heap object and may have space for more.
  3. stack ‚Äď span, which is used for goroutine stack. This span can live either in stack or in heap, but not in both.

When allocation happens we map objects into 3 size classes: Tiny class for objects <16 bytes, Small class for objects up to 32 kB and Large class for other objects. Small allocation sizes are rounded to one of about 70 size classes, each of which has its own free set of objects of exactly that size. I found some interesting comments in runtime/malloc.go about Tiny allocator, and why it was introduced: > The main targets of tiny allocator are small strings and standalone escaping variables.

On a json benchmark the allocator reduces number of allocations by ~12% and reduces heap size by ~20%.
Tiny allocator combines several tiny allocation requests into a single memory block of 16 bytes.
The resulting memory block is freed when all subobjects are unreachable.
The subobjects can’t have pointers.

So here is how it works for Tiny objects:

When allocating Tiny object:

  1. Look in to corresponding tiny slot object in this P’s mcache.
  2. Round the size of existing subobject (if exists) into 8, 4 or 2 bytes based on the new object’s size.
  3. If the object fits together with existing subobjects, place it there.

If it doesn’t fit in the tiny block:

  1. Look in the corresponding mspan in this P’s mcache.
  2. Obtain a new mspan from mcache.
  3. Scan the mspan‘s free bitmap to find a free slot.
  4. If there is a free slot, allocate it and use it as a new tiny slot object. (This can all be done without acquiring a lock.)

If the mspan has no free slots:

  1. Obtain a new mspan from the mcentral‘s list of mspans of the required size class that have free space.
  2. Obtaining a whole span amortizes the cost of locking the mcentral.

If the mspan’s list is empty:

  1. Obtain a run of pages from the mheap to use for the mspan.

If the mheap is empty or has no page runs large enough:

  1. Allocate a new group of pages (at least 1MB) from the OS.
  2. Allocating a large run of pages amortizes the cost of talking to the OS.

For Small objects it’s very similiar, but we skip the first part:

When allocating small object:

  1. Round the size up to one of the small size classes.
  2. Look in the corresponding mspan in this P’s mcache.
  3. Scan the mspan‘s free bitmap to find a free slot.
  4. If there is a free slot, allocate it. (This can all be done without acquiring a lock.)

If the mspan has no free slots:

  1. Obtain a new mspan from the mcentral‘s list of mspans of the required size class that have free space.
  2. Obtaining a whole span amortizes the cost of locking the mcentral.

If the mspan’s list is empty:

  1. Obtain a run of pages from the mheap to use for the mspan.

If the mheap is empty or has no page runs large enough:

  1. Allocate a new group of pages (at least 1MB) from the OS.
  2. Allocating a large run of pages amortizes the cost of talking to the OS.

Allocating and freeing a large object uses the mheap directly, bypassing the mcache and mcentral. mheap is managed like in TCMalloc, where we have an array of free lists. Large objects are rounded up to page size (8K) and we look for kth entry in a free list, which consists of k pages, if it’s empty we go down. Rinse and repeat, until the 128 array entry. If we don’t find empty page in 127, we look for a span in leftover large pages (mspan.freelarge field) and if that fails we take from OS.

So that’s all about Go Memory allocation, after digging thru this code runtime.MemStats makes way more sense to me. You can see all the reporting for size classes, you can view how many bytes object’s that implement memory management (like MCache, MSpan) take, etc. You can read more about memstats in exploring Prometheus Go metrics.

Getting back to the problem

So to remind you, we are investigating:

func main() {
http.HandleFunc("/bar", func(w http.ResponseWriter, r *http.Request) {
fmt.Fprintf(w, "Hello, %q", html.EscapeString(r.URL.Path))

http.ListenAndServe(":8080", nil)
go build main.go
ps -u --pid 16609
povilasv 16609 0.0 0.0 388496 5236 pts/9 Sl+ 17:21 0:00 ./main

Which gives us ~380 MiB of virtual memory size.

So maybe it’s runtime?

Let’s read memstats:

func main() {
http.HandleFunc("/bar", func(w http.ResponseWriter, r *http.Request) {
fmt.Fprintf(w, "Hello, %q", html.EscapeString(r.URL.Path))

go func() {
for {
var m runtime.MemStats

log.Println(float64(m.Sys) / 1024 / 1024)
log.Println(float64(m.HeapAlloc) / 1024 / 1024)
time.Sleep(10 * time.Second)
http.ListenAndServe(":8080", nil)

MemStats.Sys is the total bytes of memory obtained from the OS. Sys measures the virtual address space reserved by the Go runtime for the heap, stacks, and other internal data structures.
MemStats.HeapAlloc is bytes of allocated for heap objects.

Nope, it doesn’t look like it:

2018/05/08 18:00:34 4.064689636230469
2018/05/08 18:00:34 0.5109481811523438

Maybe it’s normal?

Let’s try this C program:

#include /* printf, scanf, NULL */

int main (){
int i,n;
printf ("Enter a number:");
scanf ("%d", &i);

return 0;
gcc main.c

Nope, ~10MiB:

ps -u --pid 25074

povilasv 25074 0.0 0.0 10832 908 pts/6 S+ 17:48 0:00 ./a.out

Let’s try looking into /proc

cat /proc/30376/status
Name: main
State: S (sleeping)
Pid: 30376
FDSize: 64
VmPeak: 386576 kB
VmSize: 386576 kB
VmLck: 0 kB
VmPin: 0 kB
VmHWM: 5116 kB
VmRSS: 5116 kB
RssAnon: 972 kB
RssFile: 4144 kB
RssShmem: 0 kB
VmData: 44936 kB
VmStk: 136 kB
VmExe: 2104 kB
VmLib: 2252 kB
VmPTE: 132 kB
VmSwap: 0 kB
HugetlbPages: 0 kB
CoreDumping: 0
Threads: 6

Doesn’t help as segment sizes are normal, only VmSize is huge.

Let’s take a look at /proc/maps

cat /proc/31086/maps
00400000-0060e000 r-xp 00000000 fd:01 1217120 /main
0060e000-007e5000 r--p 0020e000 fd:01 1217120 /main
007e5000-0081b000 rw-p 003e5000 fd:01 1217120 /main
0081b000-0083d000 rw-p 00000000 00:00 0
0275d000-0277e000 rw-p 00000000 00:00 0 [heap]
c000000000-c000001000 rw-p 00000000 00:00 0
c41fff0000-c420200000 rw-p 00000000 00:00 0
7face8000000-7face8021000 rw-p 00000000 00:00 0
7face8021000-7facec000000 ---p 00000000 00:00 0
7facec000000-7facec021000 rw-p 00000000 00:00 0
7facf4021000-7facf8000000 ---p 00000000 00:00 0
7facf8000000-7facf8021000 rw-p 00000000 00:00 0
7facf8021000-7facfc000000 ---p 00000000 00:00 0
7facfd323000-7facfd324000 ---p 00000000 00:00 0
7facfd324000-7facfdb24000 rw-p 00000000 00:00 0
7facfdb24000-7facfdb25000 ---p 00000000 00:00 0
7facfeb27000-7facff327000 rw-p 00000000 00:00 0
7facff327000-7facff328000 ---p 00000000 00:00 0
7facff328000-7facffb28000 rw-p 00000000 00:00 0
7fddc2798000-7fddc2f98000 rw-p 00000000 00:00 0
7fddc2f98000-7fddc2f9b000 r-xp 00000000 fd:01 2363785 libdl-2.27.so
7fddc319c000-7fddc3383000 r-xp 00000000 fd:01 2363779 libc-2.27.so
7fddc3587000-7fddc3589000 rw-p 001eb000 fd:01 2363779 libc-2.27.so
7fddc3589000-7fddc358d000 rw-p 00000000 00:00 0
7fddc358d000-7fddc35a7000 r-xp 00000000 fd:01 2363826 libpthread-2.27.so
7fddc37a8000-7fddc37ac000 rw-p 00000000 00:00 0
7fddc37ac000-7fddc37b2000 r-xp 00000000 fd:01 724559 libgtk3-nocsd.so.0
7fddc39b2000-7fddc39b3000 rw-p 00006000 fd:01 724559 libgtk3-nocsd.so.0
7fddc39b3000-7fddc39da000 r-xp 00000000 fd:01 2363771 ld-2.27.so
7fddc3af4000-7fddc3bb8000 rw-p 00000000 00:00 0
7fddc3bda000-7fddc3bdb000 r--p 00027000 fd:01 2363771 ld-2.27.so
7fddc3bdc000-7fddc3bdd000 rw-p 00000000 00:00 0
7fff472e9000-7fff4730b000 rw-p 00000000 00:00 0 [stack]
7fff473a7000-7fff473aa000 r--p 00000000 00:00 0 [vvar]
7fff473aa000-7fff473ac000 r-xp 00000000 00:00 0 [vdso]
ffffffffff600000-ffffffffff601000 r-xp 00000000 00:00 [vsyscall]

Summing all those addresses will probably leave me with the same ~380Mib. I am too lazy to sum it up. But it’s cool, scroll to the right you will see libc and other shared libraries mapped to your process.

Let’s try a simpler program

func main() {
go func() {
for {
var m runtime.MemStats

log.Println(float64(m.Sys) / 1024 / 1024)
log.Println(float64(m.HeapAlloc) / 1024 / 1024)
time.Sleep(10 * time.Second)

time.Sleep(1 * time.Hour)
go build main.go
ps -u --pid 3642
povilasv 3642 0.0 0.0 4900 948 pts/10 Sl+ 09:07 0:00 ./main

Hmm, that’s interesting, it’s only ~4Mib.

To be continued …

Thank you for reading this. As always I look forward to your comments. And please don’t spoil my search in the comments ūüėÄ


  • https://samypesse.gitbooks.io/how-to-create-an-operating-system/content/Chapter-8/
  • https://chortle.ccsu.edu/AssemblyTutorial/Chapter-04/ass04_1.html
  • http://brokenthorn.com/Resources/OSDev18.html
  • Libc man pages on memory: https://www.gnu.org/software/libc/manual/html_node/Memory.html#Memory
  • TcMalloc: http://goog-perftools.sourceforge.net/doc/tcmalloc.html
  • https://mcuoneclipse.com/2013/04/14/text-data-and-bss-code-and-data-size-explained/
  • https://stackoverflow.com/questions/24973973/is-exe-made-of-pure-machine-code-only
  • http://brokenthorn.com/Resources/OSDev18.html
  • https://www.win.tue.nl/~aeb/linux/lk/lk.html
  • https://samypesse.gitbooks.io/how-to-create-an-operating-system/content/chapter9/
  • https://wiki.osdev.org/ELF_Tutorial
  • https://stackoverflow.com/questions/610682/do-bss-section-zero-initialized-variables-occupy-space-in-elf-file
  1. Actually saying that each memory slot holds 8 bits is not really true, as there are architectures, where you can store less or more than 8 bits, you can read more in https://www.reddit.com/r/askscience/comments/3b6lkz/why_is_it_that_the_de_facto_standard_for_the/ and in https://softwareengineering.stackexchange.com/questions/91230/addressable-memory-unit#91263
  2. Actually there is a concept called shared memory, so that multiple applications can access the same memory. Read more in http://www.csl.mtu.edu/cs4411.ck/www/NOTES/process/shm/what-is-shm.html
  3. https://littleosbook.github.io/#a-short-introduction-to-virtual-memory and https://en.wikipedia.org/wiki/Virtual_memory
  4. You can read more about it in https://en.wikipedia.org/wiki/Memory_management_unit and https://wiki.osdev.org/Paging
  5. Libc man pages are a really great read: http://www.gnu.org/software/libc/manual/html_node/Memory.html
  6. The steps are for Linux, taken from https://en.wikipedia.org/wiki/Loader_(computing)
  7. You can read more about executable files in: https://en.wikipedia.org/wiki/Portable_Executable, https://wiki.osdev.org/ELF_Tutorial, https://www.quora.com/Where-does-elf-file-sits-inside-a-microcontroller
  8. You can verify that your program is in ELF, by opening the binary in text editor and seeing text ELF string in the beginning
  9. https://www.gnu.org/software/libc/manual/html_node/Memory-Concepts.html#Memory-Concepts
  10. Function definitions are in https://github.com/golang/go/blob/master/src/runtime/mem_linux.go and assembly in https://github.com/golang/go/blob/master/src/runtime/sys_linux_amd64.s#L449. There is really cool post about a bug, which happened because Go does not use libc wrappers: https://marcan.st/2017/12/debugging-an-evil-go-runtime-bug/.
  11. Read about tcmalloc design in http://goog-perftools.sourceforge.net/doc/tcmalloc.html

Exploring Prometheus Go client metrics

In this post I want to explore go metrics, which are exported by client_golang via promhttp.Handler() call.
Here is a sample program, registering prom handler and listening on 8080 port:

package main

import (


func main() {
    http.Handle("/metrics", promhttp.Handler())
    log.Fatal(http.ListenAndServe(":8080", nil))

When you hit your metrics endpoint, you will get something like:

# HELP go_gc_duration_seconds A summary of the GC invocation durations.
# TYPE go_gc_duration_seconds summary
go_gc_duration_seconds{quantile="0"} 3.5101e-05
# HELP go_goroutines Number of goroutines that currently exist.
# TYPE go_goroutines gauge
go_goroutines 6
process_open_fds 12
# HELP process_resident_memory_bytes Resident memory size in bytes.
# TYPE process_resident_memory_bytes gauge
process_resident_memory_bytes 1.1272192e+07
# HELP process_virtual_memory_bytes Virtual memory size in bytes.
# TYPE process_virtual_memory_bytes gauge
process_virtual_memory_bytes 4.74484736e+08

On initialisation client_golang registers 2 Prometheus collectors:

  • Process Collector – which collects basic Linux process information like CPU, memory,¬†file descriptor usage and start time.
  • Go Collector – which collects information about Go’s runtime like details about GC, number of gouroutines and OS threads.

Process Collector

What this collector does is reads proc file system. proc file system exposes internal kernel data structures, which is used to obtain information about the system.1

So Prometheus client reads /proc/PID/stat file, which looks like this:

1 (sh) S 0 1 1 34816 8 4194560 674 43 9 1 5 0 0 0 20 0 1 0 89724 1581056 209 18446744073709551615 94672542621696 94672543427732 140730737801568 0 0 0 0 2637828 65538 1 0 0 17 3 0 0 0 0 0 94672545527192 94672545542787 94672557428736 140730737807231 140730737807234 140730737807234 140730737807344 0

You can get human readable variant of this information using cat /proc/PID/status.

process_cpu_seconds_total – it uses utime – number of ticks executing code in user mode, measured in jiffies, with stime – jiffies spent in the system mode, executing code on behalf of the process (like doing system calls). A jiffy is the time between two ticks of the system timer interrupt. 2

process_cpu_seconds_total equals to sum of utime and stime and divide by USER_HZ. This makes sense, as dividing number of scheduler ticks by Hz(ticks per second) produces total time in seconds operating system has been running the process. 3

process_virtual_memory_bytes – uses vsize – virtual memory size is the amount of address space that a process is managing. This includes all types of memory, both in RAM and swapped out.

process_resident_memory_bytes – multiplies rss – resident set memory size is number of memory pages the process has in real memory, with pagesize 4. This results in the amount of memory that belongs specifically to that process in bytes. This excludes swapped out memory pages.

process_start_time_seconds – uses start_time – time the process started after system boot, which is expressed in jiffies and btime from /proc/stat which shows the time at which the system booted in seconds since the Unix epoch. start_time is divided by USER_HZ in order to get the value in seconds.

process_open_fds – counts the number of files in /proc/PID/fd directory. This shows currently open regular files, sockets, pseudo terminals, etc.

process_max_fds – reads /proc/{PID}/limits and uses Soft Limit from “Max Open Files” row. The interesting bit here is that /limits lists Soft and Hard limits.
As it turns out, the soft limit is the value that the kernel enforces for the corresponding resource and the Hard limit acts as a ceiling for the soft limit.
An unprivileged process may only set its soft limit to a value up to the hard limit and (irreversibly) lower its Hard limit. 5

In Go you can use err = syscall.Setrlimit(syscall.RLIMIT_NOFILE, &syscall.Rlimit{Cur: 9, Max: 10}) to set your limit.

Go Collector

Go Collector’s most of the metrics are taken from runtime, runtime/debug packages.

go_goroutines – calls out to runtime.NumGoroutine(), which computes the value based off the scheduler struct and global allglen variable. As all the values in sched struct can be changed concurently there is this funny check where if computed value is less than 1 it becomes 1.

go_threads – calls out to runtime.CreateThreadProfile(), which reads off global allm variable. If you don’t know what M’s or G’s you can read my blogpost about it.

go_gc_duration_seconds – calls out to debug.ReadGCStats() with PauseQuantile set to 5, which returns us the minimum, 25%, 50%, 75%, and maximum pause times. Then it manualy creates a Summary type from pause quantiles, NumGC var and PauseTotal seconds. It’s cool how well GCStats struct fits the prom’s Summary type. 6

go_info – this provides us with Go version. It’s pretty clever, it calls out to runtime.Version() and set’s that as a version label and then always returns value of 1 for this gauge metric.


Go Collector provides us with a lot of metrics about memory and GC.
All those metrics are from runtime.ReadMemStats(), which gives us metrics from MemStats struct.
One thing that worries me, is that runtime.ReadMemStats() has a explicit call to make a stop-the-world pause 7.
So I wonder how much actual cost this pause introduces?
As during stop-the-world pause, all goroutines are paused, so that GC can run. I’ll probably do a comparison of an app with and without instrumentation in a later post.

We already seen that Linux provides us with rss/vsize metrics for memory stats, so naturally the question arises which metrics to use, the ones provided in MemStats or rss/vsize?

The good part about resident set size and virtual memory size is that it’s based off Linux primitives and is programming language agnostic.
So in theory you could instrument any program and you would know how much memory it consumes (as long as you name your metrics consistently, ie process_virtual_memory_bytes and process_resident_memory_bytes.).
In practice, however when Go process starts up it takes a lot of virtual memory beforehand, such a simple program like the one above takes up to 544MiB of vsize on my machine (x86_64 Ubuntu), which is a bit confusing. RSS shows around 7mib, which is closer to the actual usage.

On the other hand using Go runtime based metrics gives more fined grained information on what is happening in your running application.
You should be able to find out more easily whether your program has a memory leak, how long GC took, how much it reclaimed.
Also, it should point you into right direction when you are optimizing program’s memory allocations.

I haven’t looked in detail how Go GC and memory model works, a part of it’s concurrency model 8, so this bit is still new to me.

So let’s take a look at those metrics:

go_memstats_alloc_bytes – a metric which shows how much bytes of memory is allocated on the Heap for the Objects. The value is same as go_memstats_heap_alloc_bytes. This metric counts all reachable heap objects plus unreachable objects, GC has not yet freed.

go_memstats_alloc_bytes_total – this metric increases as objects are allocated in the Heap, but doesn‚Äôt decrease when they are freed. I think it is immensly useful, as it is only increasing number and has same nice properties that Prometheus Counter has. Doing rate() on it should show us how many bytes/s of memory app consumes and is ‚Äúdurable‚ÄĚ across restarts and scrape misses.

go_memstats_sys_bytes – it’s a metric, which measures how many bytes of memory in total is taken from system by Go. It sums all the sys metrics described below.

go_memstats_lookups_total – counts how many pointer dereferences happened. This is a counter value so you can use rate() to lookups/s.

go_memstats_mallocs_total – shows how many heap objects are allocated. This is a counter value so you can use rate() to objects allocated/s.

go_memstats_frees_total – shows how many heap objects are freed. This is a counter value so you can use rate() to objects allocated/s. Note you can get number of live objects with go_memstats_mallocs_totalgo_memstats_frees_total.

Turns out that Go organizes memory in spans, which are contiguous regions of memory of 8K or larger. There are 3 types of Spans:
1) idle – span, that has no objects and can be released back to the OS, or reused for heap allocation, or reused for stack memory.
2) in use – span, that has atleast one heap object and may have space for more.
3) stack – span, which is used for goroutine stack. This span can live either in stack or in heap, but not in both.

Heap memory metrics

go_memstats_heap_alloc_bytes – same as go_memstats_alloc_bytes.

go_memstats_heap_sys_bytes – bytes of memory obtained for the heap from OS. This includes virtual address space that has been resevered, but not yet used and virtual address space which was returned to OS after it became unused. This metric estimates the largest size the heap.

go_memstats_heap_idle_bytes – shows how many bytes are in idle spans.

go_memstats_heap_idle_bytes minus go_memstats_heap_released_bytes estimates how many bytes of memory could be released, but is kept by runtime, so that runtime can allocate objects on the heap without asking OS for more memory.

go_memstats_heap_inuse_bytes – shows how many bytes in in-use spans.

go_memstats_heap_inuse_bytes minus go_memstats_heap_alloc_bytes shows how many bytes of memory has been allocated for the heap, but not currently used.

go_memstats_heap_released_bytes – shows how many bytes of idle spans were returned to the OS.

go_memstats_heap_objects – shows how many objects are allocated on the heap. This changes as GC is performed and new objects are allocated.

Stack memory metrics

go_memstats_stack_inuse_bytes – shows how many bytes of memory is used by stack memory spans, which have atleast one object in them. Go doc says, that stack memory spans can only be used for other stack spans, i.e. there is no mixing of heap objects and stack objects in one memory span.

go_memstats_stack_sys_bytes – shows how many bytes of stack memory is obtained from OS. It’s go_memstats_stack_inuse_bytes plus any memory obtained for OS thread stack.

There is no go_memstats_stack_idle_bytes, as unused stack spans are counted towards go_memstats_heap_idle_bytes.

Off-heap memory metrics

These metrics are bytes allocated for runtime internal structures, that are not allocated on the heap, because they implement the heap.

go_memstats_mspan_inuse_bytes – shows how many bytes are in use by mspan structures.

go_memstats_mspan_sys_bytes – shows how many bytes are obtained from OS for mspan structures.

go_memstats_mcache_inuse_bytes – shows how many bytes are in use by mcache structures.

go_memstats_mcache_sys_bytes – shows how many bytes are obtained from OS for mcache structures.

go_memstats_buck_hash_sys_bytes – shows how many bytes of memory are in bucket hash tables, which are used for profiling.

go_memstats_gc_sys_bytes – shows how many in garbage collection metadata.

go_memstats_other_sys_bytesgo_memstats_other_sys_bytes shows how many bytes of memory are used for other runtime allocations.

go_memstats_next_gc_bytes – shows the target heap size of the next GC cycle. GC’s goal is to keep go_memstats_heap_alloc_bytes less than this value.

go_memstats_last_gc_time_seconds – contains unix timestamp when last GC finished.

go_memstats_last_gc_cpu_fraction – shows the fraction of this program’s available CPU time used by GC since the program started.
This metric is also provided in GODEBUG=gctrace=1.

Playing around with numbers

So it’s a lot of metrics and a lot of information.
I think the best way to learn is to just play around with it, so in this part I’ll do just that.
So I’ll be using the same program that is above.
Here are the dump from /metrics (edited for space), which I’m going to use:

process_resident_memory_bytes 1.09568e+07

process_virtual_memory_bytes 6.46668288e+08

go_memstats_heap_alloc_bytes 2.24344e+06

go_memstats_heap_idle_bytes 6.3643648e+07

go_memstats_heap_inuse_bytes 3.039232e+06

go_memstats_heap_objects 6498

go_memstats_heap_released_bytes 0

go_memstats_heap_sys_bytes 6.668288e+07

go_memstats_lookups_total 0

go_memstats_frees_total 12209

go_memstats_mallocs_total 18707

go_memstats_buck_hash_sys_bytes 1.443899e+06

go_memstats_mcache_inuse_bytes 6912

go_memstats_mcache_sys_bytes 16384

go_memstats_mspan_inuse_bytes 25840

go_memstats_mspan_sys_bytes 32768

go_memstats_other_sys_bytes 1.310909e+06

go_memstats_stack_inuse_bytes 425984

go_memstats_stack_sys_bytes 425984

go_memstats_sys_bytes 7.2284408e+07

go_memstats_next_gc_bytes 4.194304e+06

go_memstats_gc_cpu_fraction 1.421928536233557e-06

go_memstats_gc_sys_bytes 2.371584e+06

go_memstats_last_gc_time_seconds 1.5235057190167596e+09

rss = 1.09568e+07 = 10956800 bytes = 10700 KiB = 10.4 MiB

vsize = 6.46668288e+08 = 646668288 bytes = 631512 KiB = 616.7 MiB

heap_alloc_bytes = 2.24344e+06 = 2243440 = 2190 KiB = 2.1 MiB

heap_inuse_bytes = 3.039232e+06 = 3039232 = 2968 KiB = 2,9 MiB

heap_idle_bytes = 6.3643648e+07 = 63643648 = 62152 KiB = 60.6 MiB

heap_released_bytes = 0

heap_sys_bytes = 6.668288e+07 = 66682880 = 65120 KiB = 63.6 MiB

frees_total = 12209

mallocs_total = 18707

mspan_inuse_bytes = 25840 = 25.2 KiB

mspan_sys_bytes = 32768 = 32 KiB

mcache_inuse_bytes = 6912 = 6.8 KiB

mcache_sys_bytes = 16384 = 12 KiB

buck_hash_sys_bytes = 1.443899e+06 = 1443899 = 1410 KiB = 1.4 MiB

gc_sys_bytes = 2.371584e+06 = 2371584 = 2316 KiB = 2.3 MiB

other_sys_bytes = 1.310909e+06 = 1310909 = 1280,2 KiB = 1.3MiB

stack_inuse_bytes = 425984 = 416 KiB

stack_sys_bytes = 425984 = 416 KiB

sys_bytes = 7.2284408e+07 = 72284408 = 70590.2 KiB = 68.9 MiB

next_gc_bytes = 4.194304e+06 = 4194304 = 4096 KiB = 4 MiB

gc_cpu_fraction = 1.421928536233557e-06 = 0.000001

last_gc_time_seconds = 1.5235057190167596e+09 = Thu, 12 Apr 2018 05:47:59 GMT

Interesting bit is that heap_inuse_bytes is more than heap_alloc_bytes.
I think heap_alloc_bytes shows how many bytes are in terms of objects and heap_inuse_bytes shows bytes of memory in terms of spans.
Dividing heap_inuse_bytes by size of the span gives: 3039232 / 8192 = 371 span.

heap_inuse_bytes minus heap_alloc_bytes, should show the amount of free space that we have in in-use spans, which is 2,9 MiB – 2.1 MiB = 0.8 MiB.
This roughly means that we can allocate 0.8 MiB of objects on the heap without using new memory spans.
But we should keep in mind memory fragmentation.
Imagine if you have a new bytes slice of 10K bytes, the memory could be in the position, where it doesn’t have a contiguous block of 10K bytes + slice header, so it would need use a new span, instead of reusing

heap_idle_bytes minus heap_released_byte shows that we have around 60.6 MiB of unused spans, which are reserved from OS and could be returned to OS. It’s 63643648/8192 = 7769 spans.

heap_sys_bytes, which is 63.6MiB estimates the largest size the heap has had. It’s 66682880/8192 = 8140 spans.

mallocs_total shows that we allocated 18707 objects and freed 12209. So currently, we have 18707-12209 = 6498 objects. We can find the average size of the object dividing heap_alloc_bytes over live objects, which is 6498. The result is 2243440 / 6498 = 345.3 bytes.
(This is probably a stupid metric, as object vary a lot in size and we should do histograms instead)

So sys_bytes should be a sum of all *sys metrics. So let’s check that:
sys_bytes == mspan_sys_bytes + mcache_sys_bytes + buck_hash_sys_bytes + gc_sys_bytes + other_sys_bytes + stack_sys_bytes + heap_sys_bytes.
So, we have 72284408 == 32768 + 16384 + 1443899 + 2371584 + 1310909 + 425984 + 66682880, which is 72284408 == 72284408, which is correct.

The interesting detail about sys_bytes, is that it’s 68,9 MiB it’s how many bytes of memory in total taken from OS. Meanwhile, OS’es vsize gives you 616,7MiB and rss 10.4 MiB. So all these numbers don’t really match up.

As I understand it so part of our memory could be in OS’s memory pages which are in swap or in filesystem (not in RAM), so this would explain why rss is smaller than sys_bytes.

And vsize contains a lot of things, like mapped libc, pthreads libs, etc. You can explore /proc/PID/maps and /proc/PID/smaps file, to see what is being currently mapped.

gc_cpu_fraction is running crazy low, 0.000001 of CPU time is used for GC. That’s really really cool. (Although this program doesn’t produce much garbage)

next_gc_bytes shows that the target for GC is to keep heap_alloc_bytes under 4 MiB, as heap_alloc_bytes is currently at 2.1 MiB the target is achieved.


I love Go and the fact that it exposes so much useful information in it’s packages and simple users like you and me can just call a function and get the information.

It was really cool playing around and reading about linux & Go, so I’m thinking of doing part 2 of this post. Maybe go look into metrics provided by cAdvisor or show how to use some of the metrics described here in dashboards/alerts with Prometheus.

Also, once vgo get’s integrated (and I really really hope it does, cause it’s like the best package manager I ever used). Then we should be able to inspect dependencies from some go runtime package, which would be really cool! Imagine writing a custom prom collector, which would go through all your dependencies, check for new versions and if found wouldgive you back a number of outdated pkgs, something like go_num_outdated_pkgs metric.
This way you could write an alert if your service get’s terribly outdated. Or check that your live dependency hashes don’t match current hashes?

If you like the post, hit the up arrow button on the reddit and see you soon.

  1. You can read more about what data kernel exposes http://tldp.org/LDP/Linux-Filesystem-Hierarchy/html/proc.html or more technical version here: https://www.kernel.org/doc/Documentation/filesystems/proc.txt
  2. Take a look at http://man7.org/linux/man-pages/man7/time.7.html, which has a great description for jiffies, user time and system time.
  3. The formula is in https://github.com/prometheus/procfs/blob/master/proc_stat.go#L187 . Also, there is an interesting work around in prometheus client definition of USER_HZ in https://github.com/prometheus/procfs/blob/master/proc_stat.go#L10-L25 .
  4. http://man7.org/linux/man-pages/man2/getpagesize.2.html
  5. More docs in https://linux.die.net/man/2/setrlimit
  6. https://golang.org/pkg/runtime/debug/#ReadGCStats
  7. https://github.com/golang/go/blob/master/src/runtime/mstats.go#L458
  8. Weirdly article “The Go Memory Model” talks about accessing memory from different goroutines https://golang.org/ref/mem rather than how memory is allocated etc

Tracking request duration with Prometheus

I recently started using Prometheus for instrumenting and I really like it! It has a cool concept of labels,  a functional query language & a bunch of very useful functions like rate(), increase() & histogram_quantile().

Adding metrics to your app is easy,  just import prometheus client and register metrics HTTP handler http.Handle("/metrics", promhttp.Handler())

This one-liner adds HTTP /metrics endpoint to HTTP router. By default client exports memory usage, number of goroutines, gc information and other runtime information. Exporting metrics as HTTP endpoint makes the whole dev/test lifecycle easy, as it is really trivial to check whether your newly added metric is now exposed.

One thing I struggled on is how to track request duration.

Prometheus doesn’t have a built in Timer metric type, which is often available in other monitoring systems. It has only 4 metric types: Counter, Gauge, Histogram and Summary.

It turns out that client library allows you to create a timer using: prometheus.NewTimer(o Observer) and record duration using ObserveDuration() method. Provided Observer can be either Summary, Histogram or a Gauge. So, which one to use?


In Prometheus Histogram is really a cumulative histogram (cumulative frequency). Histogram is made of a counter, which counts number of events that happened, a counter for a sum of event values and another counter for each of a bucket. Buckets count how many times event value was less than or equal to the bucket’s value.

Let’s take a look at the example:

Imagine that you create a histogram with 5 buckets with values:¬†0.5, 1, 2, 3, 5. Let’s call this histogram¬†http_request_duration_seconds¬†and 3 requests come in with durations 1s, 2s, 3s. Then you would see that /metrics¬†endpoint contains:

# HELP http_request_duration_seconds request duration histogram
# TYPE http_request_duration_seconds histogram
http_request_duration_seconds_bucket{le="0.5"} 0
http_request_duration_seconds_bucket{le="1"} 1
http_request_duration_seconds_bucket{le="2"} 2
http_request_duration_seconds_bucket{le="3"} 3
http_request_duration_seconds_bucket{le="5"} 3
http_request_duration_seconds_bucket{le="+Inf"} 3
http_request_duration_seconds_sum 6
http_request_duration_seconds_count 3

Here we can see that:

sum is 1s + 2s + 3s = 6,

count is 3, because of 3 requests

bucket {le=”0.5″} ¬†is 0, because none of the requests where <= 0.5 seconds

bucket {le=”1″} ¬†is 1, because one of the requests where <= 1¬†seconds

bucket {le=”2″} ¬†is 2, because two of the requests where <= 2¬†seconds

bucket {le=”3″} ¬†is 3, because all of the requests where <= 3¬†seconds

So what can we do with it?

Thirst thing to note is that when using Histogram we don’t need to have a separate counter to count total HTTP requests, as it creates one for us.

We could calculate average request time by dividing sum over count. In PromQL it would be:

http_request_duration_seconds_sum / http_request_duration_seconds_count

Also we could calculate percentiles from it. Prometheus comes with a handy histogram_quantile function for it. For example calculating 50% percentile (second quartile) for last 10 minutes in PromQL would be:

histogram_quantile(0.5, rate(http_request_duration_seconds_bucket[10m])

Which results in 1.5.

Wait, 1.5? Shouldn’t it be 2?¬†(50th percentile is supposed to be the median, the number in the middle)

I even computed the 50th percentile using cumulative frequency table (what I thought prometheus is doing) and still ended up with 2.

As it turns out, this value is only an approximation of computed quantile. You can find more information on what type of approximations prometheus is doing in histogram_quantile doc.

By the way, be warned that percentiles can be easily misinterpreted.

It’s important to understand that creating a new histogram requires you to specify bucket boundaries up front. The default values, which are 0.005, 0.01, 0.025, 0.05, 0.1, 0.25, 0.5, 1, 2.5, 5, 10¬†are tailored to broadly measure the response time in seconds and probably won’t fit your app’s behavior.

This creates a bit of a chicken or the egg problem, because you cannot know bucket boundaries until you launched the app and collected latency data and you cannot make a new Histogram without specifying (implicitly or explicitly) the bucket values. So I guess the best way to move forward is launch your app with default bucket boundaries, let it spin for a while and later tune those values based on what you see.


A Summary is like a histogram_quantile() function, but percentiles are computed in the client. Summary is made of a count and sum counters (like in Histogram type) and resulting quantile values.

For example, we want to find 0.5, 0.9, 0.99 quantiles and the same 3 requests with 1s, 2s, 3s durations come in.

Then /metrics endpoint should contain:

http_request_duration_seconds{quantile="0.5"} 2
http_request_duration_seconds{quantile="0.9"} 3
http_request_duration_seconds{quantile="0.99"} 3
http_request_duration_seconds_sum 6
http_request_duration_seconds_count 3

So we can see that:

sum is 1s + 2s + 3s = 6,

count is 3, because of 3 requests.

{quantile=”0.5″} is 2, meaning 50th percentile is 2.

{quantile=”0.9″} is 3, meaning 90th percentile is 3.

{quantile=”0.99″} is 3, meaning 99th percentile is 3.¬†

Summaries are great if¬†you already know what quantiles you want. You just specify them in¬†SummaryOpts¬†objectives map with it’s error window. For example:

map[float64]float64{0.5: 0.05}, which will compute 50th percentile with error window of 0.05.

Of course there are a couple of other parameters you could tune (like MaxAge, AgeBuckets or BufCap), but defaults should be good enough.

Although, there are a couple of problems with this approach.

First, you really need to know what percentiles you want. Because if you want to compute a different percentile, you will have to make changes in your code.

The other problem is that you cannot aggregate Summary types, i.e. if you have more than one replica of your app running you won’t be able to compute quantiles across all of the instances.

I usually don’t really know what I want, so I prefer to use Histograms.

By the way, the default go_gc_duration_seconds, which measures how long garbage collection took is implemented using Summary type.


Although Gauge doesn’t really implement¬†Observer¬†interface, you can make it using¬†prometheus.ObserverFunc(gauge.Set).

If we had the same 3 requests with 1s, 2s, 3s durations. The /metrics would contain:

# HELP http_request_duration_seconds Duration of the last http request
# TYPE http_request_duration_seconds gauge
http_request_duration_seconds 3

So we can see that:

http_request_duration_seconds is 3, meaning that last observed duration was 3.

At first I thought, this is great, I’ll just record all my request durations this way and aggregate/average out them later.

Well, no.

The mistake here is that Prometheus scrapes /metrics data only once in a while (by default every 1 min), which is configured by scrap_interval for your target.

So if you don’t have a lot of requests you could try to configure scrape_interval¬†to align with your requests and then you would see how long each request took. But I don’t think it’s a good idea, in this case I would rather push¬†the Gauge metrics to Prometheus.

I think this could be useful¬†for “job” type problems . For example, you could push how long backup, or data aggregating job has took.


Thanks for reading. Hopefully by now you and I know a bit more about Histograms, Summaries and tracking request duration. My plan for now is to track latency using Histograms, play around with histogram_quantile  and make some beautiful dashboards.

Oh and I forgot to mention, if you are instrumenting HTTP server or client, prometheus library has some helpers around it in promhttp package.