Linux perf 示例



These are some examples of using the perf Linux profiler, which has also been called Performance Counters for Linux (PCL), Linux perf events (LPE), or perf_events. Like Vince Weaver, I’ll call it perf_events so that you can search on that term later. Searching for just “perf” finds sites on the police, petroleum, weed control, and a T-shirt. This is not an official perf page, for either perf_events or the T-shirt.


perf_events is an event-oriented observability tool, which can help you solve advanced performance and troubleshooting functions. Questions that can be answered include:

  • Why is the kernel on-CPU so much? What code-paths?
  • Which code-paths are causing CPU level 2 cache misses?
  • Are the CPUs stalled on memory I/O?
  • Which code-paths are allocating memory, and how much?
  • What is triggering TCP retransmits?
  • Is a certain kernel function being called, and how often?
  • What reasons are threads leaving the CPU?

perf_events is part of the Linux kernel, under tools/perf. While it uses many Linux tracing features, some are not yet exposed via the perf command, and need to be used via the ftrace interface instead. My perf-tools collection (github) uses both perf_events and ftrace as needed.

This page includes my examples of perf_events. A table of contents:

  1. Screenshot
2. One-Liners
3. Presentations
4. Background
4.1. Prerequisites
4.2. Symbols
4.3. JIT Symbols (Java, Node.js)
4.4. Stack Traces
4.5. Audience
4.6. Usage
4.7. Usage Examples
  5. Events
5.1. PMU Events
5.2. Tracepoints
5.3. Dynamic Tracing
6. Examples
6.1. CPU Statistics
6.2. CPU Profiling
6.3. Instruction Profiling
6.4. Static Tracing
6.5. Dynamic Tracing
6.6. BPF
  7. Visualizations
7.1. Flame Graphs
7.2. Heat Maps
8. More
9. Building
10. Troubleshooting
11. Other Tools
12. Resources

Key sections to start with are: Events, One-Liners, Presentations, Prerequisites, CPU statistics, CPU profiling, and Flame Graphs. Also see my Posts about perf_events, and Links for the main (official) perf_events page, awesome tutorial, and other links. The next sections introduce perf_events further, starting with a screenshot, one-liners, and then background.

This page is under construction, and there’s a lot more to perf_events that I’d like to add. Hopefully this is useful so far.

1. Screenshot

I like explaining tools by starting with an actual screenshot showing something useful. Here’s perf version 3.9.3 tracing disk I/O:

A perf record command was used to trace the block:block_rq_issue probe, which fires when a block device I/O request is issued (disk I/O). Options included -a to trace all CPUs, and -g to capture call graphs (stack traces). Trace data is written to a file, and tracing ended when Ctrl-C was hit. A summary of the file was printed using perf report, which builds a tree from the stack traces, coalescing common paths, and showing percentages for each path.

The perf report output shows that 2,216 events were traced (disk I/O), 32% of which from a dd command. These were issued by the kernel function blk_peek_request(), and walking down the stacks, about half of these 32% were from the close() system call.

2. One-Liners

Some useful one-liners I’ve gathered or written:

Listing Events

Counting Events


Static Tracing

Dynamic Tracing


These one-liners serve to illustrate the capabilities of perf_events, and can also be used a bite-sized tutorial: learn perf_events one line at a time. You can also print these out as a perf_events cheatsheet.

3. Presentations

SCALE13x (2015)

At the Southern California Linux Expo (SCALE 13x), I gave a talk on Linux Profiling at Netflix using perf_events (aka “perf”), covering CPU profiling and a tour of other features. This talk includes a crash course on perf_events that you may find useful, plus it covers gotchas such as fixing stack traces and symbols when profiling Java and Node.js.

A video of the talk is on youtube, and the slides are on slideshare.


In a post about this talk, I included the interactive CPU flame graph SVG I was demonstrating.

4. Background

The following sections provide some background for understanding perf_events and how to use it. I’ll describe the prerequisites, audience, usage, events, and tracepoints.

4.1. Prerequisites

The perf tool is in the linux-tools-common package. Start by adding that, then running “perf” to see if you get the USAGE message. It may tell you to install another related package (linux-tools-kernelversion).

You can also build and add perf from the Linux kernel source. See the Building section.

To get the most out perf, you’ll want symbols and stack traces. These may work by default in your Linux distribution, or they may require the addition of packages, or recompilation of the kernel with additional config options.

4.2. Symbols

perf_events, like other debug tools, needs symbol information (symbols). These are used to translate memory addresses into function and variable names, so that they can be read by us humans. Without symbols, you’ll see hexadecimal numbers representing the memory addresses profiled.

The following perf report output shows stack traces, however, only hexadecimal numbers can be seen:

If the software was added by packages, you may find debug packages (often “-dbgsym”) which provide the symbols. Sometimes perf report will tell you to install these, eg: “no symbols found in /bin/dd, maybe install a debug package?”.

Here’s the same perf report output seen earlier, after adding openssh-server-dbgsym and libc6-dbgsym (this is on ubuntu 12.04):

I find it useful to add both libc6-dbgsym and coreutils-dbgsym, to provide some symbol coverage of user-level OS codepaths.

Another way to get symbols is to compile the software yourself. For example, I just compiled node (Node.js):

This has not been stripped, so I can profile node and see more than just hex. If the result is stripped, configure your build system not to run strip(1) on the output binaries.

Kernel-level symbols are in the kernel debuginfo package, or when the kernel is compiled with CONFIG_KALLSYMS.

4.3. JIT Symbols (Java, Node.js)

Programs that have virtual machines (VMs), like Java’s JVM and node’s v8, execute their own virtual processor, which has its own way of executing functions and managing stacks. If you profile these using perf_events, you’ll see symbols for the VM engine, which have some use (eg, to identify if time is spent in GC), but you won’t see the language-level context you might be expecting. Eg, you won’t see Java classes and methods.

perf_events has JIT support to solve this, which requires the VM to maintain a /tmp/ file for symbol translation. Java can do this with perf-map-agent, and Node.js 0.11.13+ with –perf_basic_prof. See my blog post Node.js flame graphs on Linux for the steps.

Note that Java may not show full stacks to begin with, due to hotspot on x86 omitting the frame pointer (just like gcc). On newer versions (JDK 8u60+), you can use the -XX:+PreserveFramePointer option to fix this behavior, and profile fully using perf. See my Netflix Tech Blog post, Java in Flames, for a full writeup, and my Java flame graphs section, which links to an older patch and includes an example resulting flame graph.

4.4 Stack Traces

Always compile with frame pointers. Omitting frame pointers is an evil compiler optimization that breaks debuggers, and sadly, is often the default. Without them, you may see incomplete stacks from perf_events, like seen in the earlier sshd symbols example. There are two ways to fix this: either using dwarf data to unwind the stack, or returning the frame pointers.


Since about the 3.9 kernel, perf_events has supported a workaround for missing frame pointers in user-level stacks: libunwind, which uses dwarf. This can be enabled using “-g dwarf”.

Frame Pointers

The earlier sshd example was a default build of OpenSSH, which uses compiler optimizations (-O2), which in this case has omitted the frame pointer. Here’s how it looks after recompiling OpenSSH with -fno-omit-frame-pointer:

Now the ancestry from add_one_listen_addr() can be seen, down to main() and __libc_start_main().

The kernel can suffer the same problem. Here’s an example CPU profile collected on an idle server, with stack traces (-g):

The kernel stack traces are incomplete. Now a similar profile with CONFIG_FRAME_POINTER=y:

Much better — the entire path from the write() syscall (__write_nocancel) to iowrite16() can be seen.

Also see the Building section for other notes about building perf_events.

4.5. Audience

To use perf_events, you’ll either:

  • Develop your own commands
  • Run example commands

Developing new invocations of perf_events requires the study of kernel and application code, which isn’t for everyone. Many more people will use perf_events by running commands developed by other people, like the examples on this page. This can work out fine: your organization may only need one or two people who can develop perf_events commands or source them, and then share them for use by the entire operation and support groups.

Either way, you need to know the capabilities of perf_events so you know when to reach for it, whether that means searching for an example command or writing your own. One goal of the examples that follow is just to show you what can be done, to help you learn these capabilities. You should also browse examples on other sites (Links).

If you’ve never used perf_events before, you may want to test before production use (it has had kernel panic bugs in the past). My experience has been a good one (no panics).

4.6. Usage

perf_events provides a command line tool, perf, and subcommands for various profiling activities. This is a single interface for the different instrumentation frameworks that provide the various events.

The perf command alone will list the subcommands; here is perf version 3.9.3 (for the Linux 3.9.3 kernel):

Apart from separate help for each subcommand, there is also documentation in the kernel source under tools/perf/Documentation. Note that perf has evolved, with different functionality added over time, and so its usage may not feel consistent as you switch between activities. It’s best to think of it as a multi-tool.

perf_events can instrument in three ways (using the perf_events terminology):

  • counting events in-kernel context, where a summary of counts is printed by perf. This mode does not generate a file.
  • sampling events, which writes event data to a kernel buffer, which is read at a gentle asynchronous rate by the perf command to write to the file. This file is then read by the perf report or perf script commands.
  • bpf programs on events, a new feature in Linux 4.4+ kernels that can execute custom user-defined programs in kernel space, which can perform efficient filters and summaries of the data. Eg, efficiently-measured latency histograms.

Try starting by counting events using the perf stat command, to see if this is sufficient. This subcommand costs the least overhead.

When using the sampling mode with perf record, you’ll need to be a little careful about the overheads, as the capture files can quickly become hundreds of Mbytes. It depends on the rate of the event you are tracing: the more frequent, the higher the overhead and larger the size.

To really cut down overhead and generate more advanced summaries, write BPF programs executed by perf. See the BPF section.

4.7. Usage Examples

These example sequences have been chosen to illustrate some different ways that perf is used, from gathering to reporting.

Performance counter summaries, including IPC, for the gzip command:

Count all scheduler process events for 5 seconds, and count by tracepoint:

Trace all scheduler process events for 5 seconds, and count by both tracepoint and process name:

Trace all scheduler process events for 5 seconds, and dump per-event details:

Trace read() syscalls, when requested bytes is less than 10:

Sample CPU stacks at 99 Hertz, for 5 seconds:

Dynamically instrument the kernel tcp_sendmsg() function, and trace it for 5 seconds, with stack traces:

Deleting the tracepoint (–del) wasn’t necessary; I included it to show how to return the system to its original state.


The use of -p PID as a filter doesn’t work properly on some kernel versions: perf hits 100% CPU and needs to be killed. It’s annoying. The workaround is to profile all CPUs (-a), and filter PIDs later.

5. Events

perf_events instruments “events”, which are a unified interface for different kernel instrumentation frameworks. The following map (from my SCaLE13x talk) illustrates the event sources:

The types of events are:

  • Hardware Events: These instrument low-level processor activity based on CPU performance counters. For example, CPU cycles, instructions retired, memory stall cycles, level 2 cache misses, etc. Some will be listed as Hardware Cache Events.
  • Software Events: These are low level events based on kernel counters. For example, CPU migrations, minor faults, major faults, etc.
  • Tracepoint Events: This are kernel-level events based on the ftrace framework. These tracepoints are placed in interesting and logical locations of the kernel, so that higher-level behavior can be easily traced. For example, system calls, TCP events, file system I/O, disk I/O, etc. These are grouped into libraries of tracepoints; eg, “sock:” for socket events, “sched:” for CPU scheduler events.
  • Dynamic Tracing: Software can be dynamically instrumented, creating events in any location. For kernel software, this uses the kprobes framework. For user-level software, uprobes.
  • Timed Profiling: Snapshots can be collected at an arbitrary frequency, using perf record -FHz. This is commonly used for CPU usage profiling, and works by creating custom timed interrupt events.

Details about the events can be collected, including timestamps, the code path that led to it, and other specific details. The capabilities of perf_events are enormous, and you’re likely to only ever use a fraction.

Currently available events can be listed using the list subcommand:

When you use dynamic tracing, you are extending this list. The probe:tcp_sendmsg tracepoint in this list is an example, which I added by instrumenting tcp_sendmsg(). Profiling (sampling) events are not listed.

5.1. PMU Events

XXX: Intro here. Performance Monitoring Counters.

For examples of using PMCs, see 6.1. CPU Statistics.

5.2. Tracepoints

Summarizing the tracepoint library names and numbers of tracepoints, on my system:

These include:

  • block: block device I/O
  • ext3, ext4: file system operations
  • kmem: kernel memory allocation events
  • random: kernel random number generator events
  • sched: CPU scheduler events
  • syscalls: system call enter and exits
  • task: task events

It’s worth checking the list of tracepoints after every kernel upgrade, to see if any are new. The value of adding them has been debated from time to time, with it wondered if anyone will use them (I will!). There is a balance to aim for: I’d include the smallest number of probes that sufficiently covers common needs, and anything unusual or uncommon can be left to dynamic tracing.

For examples of using tracepoints, see 6.4. Static Tracing.

5.3. Dynamic Tracing

The difference between tracepoints and dynamic tracing is shown in the following figure, which illustrates the coverage of common tracepoint libraries:

While dynamic tracing can see everything, it’s also an unstable interface since it is instrumenting raw code. That means that any dynamic tracing tools you develop may break after a kernel patch or update. Try to use the static tracepoints first, since their interface should be much more stable. They can also be easier to use and understand, since they have been designed with a tracing end-user in mind.

For examples of using dynamic tracing, see 6.5. Dynamic Tracing.

6. Examples

These are some examples of perf_events, collected from a variety of 3.x Linux systems.

6.1. CPU Statistics

perf_events began life as a tool for instrumenting the processor’s performance monitoring unit (PMU) hardware counters, also called performance monitoring counters (PMCs) and performance instrumentation counters (PICs). The perf stat command instruments and summarizes key counters. This is from perf version

This includes instructions per cycle (IPC), labled “insns per cycle”, or in earlier versions, “IPC”. This is a commonly examined metric, either IPC or its invert, CPI. Higher IPC values mean higher instruction throughput, and lower values indicate more stall cycles. I’d generally interpret high IPC values (eg, over 1.0) as good, indicating optimal processing of work. However, I’d want to double check what the instructions are, in case this is due to a spin loop: a high rate of instructions, but a low rate of actual work completed.

There are some advanced metrics now included in perf stat: frontend cycles idle, backend cycles idle, and stalled cycles per insn. To really understand these, you’ll need some knowledge of CPU microarchitecture.

CPU Microarchitecture

The frontend and backend metrics refer to the CPU pipeline, and are also based on stall counts. The frontend processes CPU instructions, in order. It involves instruction fetch, along with branch prediction, and decode. The decoded instructions become micro-operations (uops) which the backend processes, and it may do so out of order. For a longer summary of these components, see Shannon Cepeda’s great posts on frontend and backend.

The backend can also process multiple uops in parallel; for modern processors, three or four. Along with pipelining, this is how IPC can become greater than one, as more than one instruction can be completed (“retired”) per CPU cycle.

Stalled cycles per instruction is similar to IPC (inverted), however, only counting stalled cycles, which will be for memory or resource bus access. This makes it easy to interpret: stalls are latency, reduce stalls. I really like it as a metric, and hope it becomes as commonplace as IPC/CPI. Lets call it SCPI.

Detailed Mode

There is a “detailed” mode for perf stat:

This includes additional counters for Level 1 data cache events, and last level cache (LLC) events.

Specific Counters

Hardware cache event counters, seen in perf list, can be instrumented. Eg:

The percentage printed is a convenient calculation that perf_events has included, based on the counters I specified. If you include the “cycles” and “instructions” counters, it will include an IPC calculation in the output.

These hardware events that can be measured are often specific to the processor model. Many may not be available from within a virtualized environment.

Raw Counters

The Intel 64 and IA-32 Architectures Software Developer’s Manual Volume 3B: System Programming Guide, Part 2 and the BIOS and Kernel Developer’s Guide (BKDG) For AMD Family 10h Processors are full of interesting counters, but most cannot be found in perf list. If you find one you want to instrument, you can specify it as a raw event with the format: rUUEE, where UU == umask, and EE == event number. Here’s an example where I’ve added a couple of raw counters:

If I did this right, then r80a2 has instrumented RESOURCE_STALLS.OTHER, and r2b1 has instrumented UOPS_DISPATCHED.CORE: the number of uops dispatched each cycle. It’s easy to mess this up, and you’ll want to double check that you are on the right page of the manual for your processor.

If you do find an awesome raw counter, please suggest it be added as an alias in perf_events, so we all can find it in perf list.

Other Options

The perf subcommands, especially perf stat, have an extensive option set which can be listed using “-h”. I’ve included the full output for perf stat here from version 3.9.3, not as a reference, but as an illustration of the interface:

Options such as –repeat, –sync, –pre, and –post can be quite useful when doing automated testing or micro-benchmarking.

6.2. CPU Profiling

perf_events can profile CPU usage based on sampling at a fixed interval. It can also instrument CPU hardware counters, and profile based on cache misses, memory stall cycles, and other low-level processor events.


Sampling CPU stacks at 99 Hertz (-F 99), for the entire system (-a, for all CPUs), with stack traces (-g, for call graphs), for 10 seconds:

The choice of 99 Hertz, instead of 100 Hertz, is to avoid accidentally sampling in lockstep with some periodic activity, which would produce skewed results. This is also coarse: you may want to increase that to higher rates (eg, up to 997 Hertz) for finer resolution, especially if you are sampling short bursts of activity and you’d still like enough resolution to be useful. Bear in mind that higher frequencies means higher overhead.

The file can be processed in a variety of ways. On recent versions, the perf report command launches an ncurses navigator for call graph inspection. Older versions of perf (or if use –stdio in the new version) print the call graph as a tree, annotated with percentages:

This tree starts with the on-CPU functions and works back through the ancestry. This approach is called a “callee based call graph”. This can be flipped by using -G for an “inverted call graph”, or by using the “caller” option to -g/–call-graph, instead of the “callee” default.

The hottest (most frequent) stack trace in this perf call graph occurred in 90.99% of samples, which is the product of the overhead percentage and top stack leaf (94.12% x 96.67%, which are relative rates). perf report can also be run with “-g graph” to show absolute overhead rates, in which case “90.99%” is directly displayed on the stack leaf:

If user-level stacks look incomplete, you can try perf record with “-g dwarf” as a different technique to unwind them. See the Stacks section.

The output from perf report can be many pages long, which can become cumbersome to read. Try generating Flame Graphs from the same data.

Hardware Counters

Apart from sampling at a fixed interval, taking samples triggered by CPU hardware counters is another form of CPU profiling. The available events can be found using perf list:

For many of these, gathering a stack on every occurrence would induce far too much overhead, and would slow down the system and change the performance characteristics of the target. It’s usually sufficient to only instrument a small fraction of their occurrences, rather than all of them. This can be done by specifying a threshold for triggering event collection, using “-c” and a count.

For example, the following one-liner instruments Level 1 data cache load misses, collecting a stack trace for one in every 10,000 occurrences:

The mechanics of “-c count” are implemented by the processor, which only interrupts the kernel when the threshold has been reached.

See the earlier Raw Counters section for an example of specifying a custom counter.

6.3. Instruction Profiling and Other Counters

XXX: fill me in. Need to cover problems and PEBS support, and other counters. There’s a lot to document when I get a chance.

6.4. Static Tracing

The following examples demonstrate static tracing: the instrumentation of tracepoints and other static events.

Counting Syscalls

The following simple one-liner counts system calls for the executed command, and prints a summary (of non-zero counts):

In this case, a gzip command was analyzed. The report shows that there were 3,201 write() syscalls, and half that number of read() syscalls. Many of the other syscalls will be due to process and library initialization.

A similar report can be seen using strace -c, the system call tracer, however it may induce much higher overhead than perf, as perf buffers data in-kernel.

perf vs strace

To explain the difference a little further: the current implementation of strace uses ptrace(2) to attach to the target process and stop it during system calls, like a debugger. This is violent, and can cause serious overhead. To demonstrate this, the following syscall-heavy program was run by itself, with perf, and with strace. I’ve only included the line of output that shows its performance:

With perf, the program ran 2.5x slower. But with strace, it ran 62x slower. That’s likely to be a worst-case result: if syscalls are not so frequent, the difference between the tools will not be as great.

Recent version of perf have included a trace subcommand, to provide some similar functionality to strace, but with much lower overhead.

New Processes

Tracing new processes triggered by a “man ls”:

Nine different commands were executed, each once. I used -n to print the “Samples” column, and “–sort comm” to customize the remaining columns.

This works by tracing sched:sched_process_exec, when a process runs exec() to execute a different binary. This is often how new processes are created, but not always. An application may fork() to create a pool of worker processes, but not exec() a different binary. An application may also reexec: call exec() again, on itself, usually to clean up its address space. In that case, it’s will be seen by this exec tracepoint, but it’s not a new process.

The sched:sched_process_fork tracepoint can be traced to only catch new processes, created via fork(). The downside is that the process identified is the parent, not the new target, as the new process has yet to exec() it’s final program.

Outbound Connections

There can be times when it’s useful to double check what network connections are initiated by a server, from which processes, and why. You might be surprised. These connections can be important to understand, as they can be a source of latency.

For this example, I have a completely idle ubuntu server, and while tracing I’ll login to it using ssh. I’m going to trace outbound connections via the connect() syscall. Given that I’m performing an inbound connection over SSH, will there be any outbound connections at all?

The report shows that sshd, groups, mesg, and bash are all performing connect() syscalls. Ring a bell?

The stack traces that led to the connect() can explain why:

Ah, these are nscd calls: the name service cache daemon. If you see hexadecimal numbers and not function names, you will need to install debug info: see the earlier section on Symbols. These nscd calls are likely triggered by calling getaddrinfo(), which server software may be using to resolve IP addresses for logging, or for matching hostnames in config files. Browsing the stack traces should identify why.

For sshd, this was called via add_one_listen_addr(): a name that was only visible after adding the openssh-server-dbgsym package. Unfortunately, the stack trace doesn’t continue after add_one_listen_add(). I can browse the OpenSSH code to figure out the reasons we’re calling into add_one_listen_add(), or, I can get the stack traces to work. See the earlier section on Stack Traces.

I took a quick look at the OpenSSH code, and it looks like this code-path is due to parsing ListenAddress from the sshd_config file, which can contain either an IP address or a hostname.

Socket Buffers

Tracing the consumption of socket buffers, and the stack traces, is one way to identify what is leading to socket or network I/O.

The swapper stack shows the network receive path, triggered by an interrupt. The sshd path shows writes.

6.5. Dynamic Tracing

For kernel analysis, I’m using CONFIG_KPROBES=y and CONFIG_KPROBE_EVENTS=y, to enable kernel dynamic tracing, and CONFIG_FRAME_POINTER=y, for frame pointer-based kernel stacks. For user-level analysis, CONFIG_UPROBES=y and CONFIG_UPROBE_EVENTS=y, for user-level dynamic tracing.

Kernel: tcp_sendmsg()

This example shows instrumenting the kernel tcp_sendmsg() function on the Linux 3.9.3 kernel:

This adds a new tracepoint event. It suggests using the -R option, to collect raw sample records, which is already the default for tracepoints. Tracing this event for 5 seconds, recording stack traces:

And the report:

This shows the path from the write() system call to tcp_sendmsg().

You can delete these dynamic tracepoints if you want after use, using perf probe --del.

Kernel: tcp_sendmsg() with size

If your kernel has debuginfo (CONFIG_DEBUG_INFO=y), you can fish out kernel variables from functions. This is a simple example of examining a size_t (integer), on Linux 3.13.1.

Listing variables available for tcp_sendmsg():

Creating a probe for tcp_sendmsg() with the “size” variable:

Tracing this probe:

The size is shown as hexadecimal.

Kernel: tcp_sendmsg() line number and local variable

With debuginfo, perf_events can create tracepoints for lines within kernel functions. Listing available line probes for tcp_sendmsg():

This is Linux 3.14.5; your kernel version may look different. Lets check what variables are available on line 81:

Now lets trace line 81, with the seglen variable that is checked in the loop:

This is pretty amazing. Remember that you can also include in-kernel filtering using –filter, to match only the data you want.

User: malloc()

While this is an interesting example, I want to say right off the bat that malloc() calls are very frequent, so you will need to consider the overheads of tracing calls like this.

Adding a libc malloc() probe:

Tracing it system-wide:

The report:

This shows the most malloc() calls were by apt-config, while I was tracing.

User: malloc() with size

As of the Linux 3.13.1 kernel, this is not supported yet:

As a workaround, you can access the registers (on Linux 3.7+). For example, on x86_64:

These registers (“%di” etc) are dependent on your processor architecture. To figure out which ones to use, see the X86 calling conventions on Wikipedia, or page 24 of the AMD64 ABI (PDF). (Thanks Jose E. Nunez for digging out these references.)

6.6. BPF

XXX: fill me in. As of Linux 4.4, perf has some enhanced BPF support (formally “eBPF”), with more in later kernels. This section should show how to write BPF programs, invoke them with perf, use cases, etc. BPF makes perf tracing programmatic, and takes perf from being a counting & sampling-with-post-processing tracer, to a fully in-kernel programmable tracer.

7. Visualizations

perf_events has a builtin visualization: timecharts, as well as text-style visualization via its text user interface (TUI) and tree reports. The following two sections show visualizations of my own: flame graphs and heat maps. The software I’m using is open source and on github, and produces these from perf_events collected data.

7.1. Flame Graphs

Flame Graphs can be produced from perf_events profiling data using the FlameGraph tools software. This visualizes the same data you see in perf report, and works with any file that was captured with stack traces (-g).


This example CPU flame graph shows a network workload for the 3.2.9-1 Linux kernel, running as a KVM instance (SVG, PNG):

Flame Graphs show the sample population across the x-axis, and stack depth on the y-axis. Each function (stack frame) is drawn as a rectangle, with the width relative to the number of samples. See the CPU Flame Graphs page for the full description of how these work.

You can use the mouse to explore where kernel CPU time is spent, quickly quantifying code-paths and determining where performance tuning efforts are best spent. This example shows that most time was spent in the vp_notify() code-path, spending 70.52% of all on-CPU samples performing iowrite16(), which is handled by the KVM hypervisor. This information has been extremely useful for directing KVM performance efforts.

A similar network workload on a bare metal Linux system looks quite different, as networking isn’t processed via the virtio-net driver, for a start.


The example flame graph was generated using perf_events and the FlameGraph tools:

The first perf command profiles CPU stacks, as explained earlier. I adjusted the rate to 99 Hertz here; I actually generated the flame graph from a 1000 Hertz profile, but I’d only use that if you had a reason to go faster, which costs more in overhead. The samples are saved in a file, which can be viewed using perf report:

This tree follows the flame graph when reading it top-down. When using -g/–call-graph (for “caller”, instead of the “callee” default), it generates a tree that follows the flame graph when read bottom-up. The hottest stack trace in the flame graph (@70.52%) can be seen in this perf call graph as the product of the top three nodes (72.18% x 99.53% x 98.16%).

The perf report tree (and the ncurses navigator) do an excellent job at presenting this information as text. However, with text there are limitations. The output often does not fit in one screen (you could say it doesn’t need to, if the bulk of the samples are identified on the first page). Also, identifying the hottest code paths requires reading the percentages. With the flame graph, all the data is on screen at once, and the hottest code-paths are immediately obvious as the widest functions.

For generating the flame graph, the perf script command dumps the stack samples, which are then aggregated by and folded into single lines per-stack. That output is then converted by into the SVG. I included a gratuitous “cat” command to make it clear that can process the output of a pipe, which could include Unix commands to filter or preprocess (grep, sed, awk).


A flame graph can be generated directly by piping all the steps:

In practice I don’t do this, as I often re-run multiple times, and this one-liner would execute everything multiple times. The output of perf script can be dozens of Mbytes, taking many seconds to process. By writing to a file, you’ve cached the slowest step, and can also edit the file (vi) to delete unimportant stacks, such as CPU idle threads.


The one-line-per-stack output of is also convenient for grep(1). Eg:

I frequently elide the cpu_idle threads in this way, to focus on the real threads that are consuming CPU resources. If I miss this step, the cpu_idle threads can often dominate the flame graph, squeezing the interesting code paths.

Note that it would be a little more efficient to process the output of perf report instead of perf script; better still, perf report could have a report style (eg, “-g folded”) that output folded stacks directly, obviating the need for There could even be a perf mode that output the SVG directly (which wouldn’t be the first one; see perf-timechart), although, that would miss the value of being able to grep the folded stacks (which I use frequently).

There are more examples of perf_events CPU flame graphs on the CPU flame graph page, including a summary of these instructions.

7.2. Heat Maps

Since perf_events can record high resolution timestamps (microseconds) for events, some latency measurements can be derived from trace data.


The following heat map visualizes disk I/O latency data collected from perf_events (SVG, PNG):

Mouse-over blocks to explore the latency distribution over time. The x-axis is the passage of time, the y-axis latency, and the z-axis (color) is the number of I/O at that time and latency range. The distribution is bimodal, with the dark line at the bottom showing that many disk I/O completed with sub-millisecond latency: cache hits. There is a cloud of disk I/O from about 3 ms to 25 ms, which would be caused by random disk I/O (and queueing). Both these modes averaged to the 9 ms we saw earlier.

The following iostat output was collected at the same time as the heat map data was collected (shows a typical one second summary):

This workload has an average I/O time (await) of 9 milliseconds, which sounds like a fairly random workload on 7200 RPM disks. The problem is that we don’t know the distribution from the iostat output, or any similar latency average. There could be latency outliers present, which is not visible in the average, and yet are causing problems. The heat map did show I/O up to 50 ms, which you might not have expected from that iostat output. There could also be multiple modes, as we saw in the heat map, which are also not visible in an average.


I used perf_events to record the block request (disk I/O) issue and completion static tracepoints:

The full output from perf script is about 70,000 lines. I’ve included some here so that you can see the kind of data available.


To calculate latency for each I/O, I’ll need to pair up the issue and completion events, so that I can calculate the timestamp delta. The columns look straightforward (and are in include/trace/events/block.h), with the 4th field the timestamp in seconds (with microsecond resolution), the 6th field the disk device ID (major, minor), and a later field (which varies based on the tracepoint) has the disk offset. I’ll use the disk device ID and offset as the unique identifier, assuming the kernel will not issue concurrent I/O to the exact same location.

I’ll use awk to do these calculations and print the completion times and latency:

I converted both columns to be microseconds, to make the next step easier.


Now I can use my program (github), to generate the interactive SVG heatmap from the trace data (and uses microseconds by default):

When I generated the heatmap, I truncated the y scale to 50 ms. You can adjust it to suit your investigation, increasing it to see more of the latency outliers, or decreasing it to reveal more resolution for the lower latencies: for example, with a 250 us limit.


While this can be useful to do, be mindful of overheads. In my case, I had a low rate of disk I/O (~300 IOPS), which generated an 8 Mbyte trace file after 2 minutes. If your disk IOPS were 100x that, your trace file will also be 100x, and the overheads for gathering and processing will add up.

For more about latency heatmaps, see my LISA 2010 presentation slides, and my CACM 2010 article, both about heat maps. Also see my Perf Heat Maps blog post.

8. More

There’s more capabilities to perf_events than I’ve demonstrated here. I’ll add examples of the other subcommands when I get a chance.

Here’s a preview of perf trace, which was added in 3.7, demonstrated on 3.13.1:

An advantage is that this is buffered tracing, which costs much less overhead than strace, as I described earlier. The perf trace output seen from this 3.13.1 kernel does, however, looks suspicious for a number of reasons. I think this is still an in-development feature. It reminds me of my dtruss tool, which has a similar role, before I added code to print each system call in a custom and appropriate way.

9. Building

The steps to build perf_events depends on your kernel version and Linux distribution. In summary:

  1. Get the Linux kernel source that matches your currently running kernel (eg, from the linux-source package, or
  2. Unpack the kernel source.
  3. cd tools/perf
  4. make
  5. Fix all errors, and most warnings, from (4).

The first error may be that you are missing make, or a compiler (gcc). Once you have those, you may then see various warnings about missing libraries, which disable perf features. I’d install as many as possible, and take note of the ones you are missing.

These perf build warnings are really helpful, and are generated by its Makefile. Here’s the makefile from 3.9.3:

Take the time to read them. This list is likely to grow as new features are added to perf_events.

The following notes show what I’ve specifically done for kernel versions and distributions, in case it is helpful.

Packages: Ubuntu, 3.8.6

Packages required for key functionality: gcc make bison flex elfutils libelf-dev libdw-dev libaudit-dev. You may also consider python-dev (for python scripting) and binutils-dev (for symbol demangling), which are larger packages.

Kernel Config: 3.8.6

Here are some kernel CONFIG options for perf_events functionality:

You may need to build your own kernel to enable these. The exact set you need depends on your needs and kernel version, and list is likely to grow as new features are added to perf_events.

10. Troubleshooting

If you see hexadecimal numbers instead of symbols, or have truncated stack traces, see the Prerequisites section.

Here are some rough notes from other issues I’ve encountered.

This sometimes works ( and sometimes throws the following error (3.9.3):

This can be fixed by increasing the file descriptor limit using ulimit -n.

Type 3 errors:

11. Other Tools

perf_events has the capabilities from many other tools rolled into one: strace(1), for tracing system calls, tcpdump(8), for tracing network packets, and blktrace(1), for tracing block device I/O (disk I/O), and other targets including file system and scheduler events. Tracing all events from one tool is not only convenient, it also allows direct correlations, including timestamps, between different instrumentation sources. Unlike these other tools, some assembly is required, which may not be for everyone (as explained in Audience).

12. Resources

Resources for further study.

12.1. Posts

I’ve been writing blog posts on specific perf_events topics. My suggested reading order is from oldest to newest (top down):

And posts on ftrace:

12.2. Links