Simpleperf

Simpleperf is a native profiling tool for Android. It can be used to profile both Android applications and native processes running on Android. It can profile both Java and C++ code on Android. It can be used on Android L and above.

Simpleperf is part of the Android Open Source Project. The source code is here. The latest document is here.

Table of Contents

Introduction

Simpleperf contains two parts: the simpleperf executable and Python scripts.

The simpleperf executable works similar to linux-tools-perf, but has some specific features for the Android profiling environment:

  1. It collects more info in profiling data. Since the common workflow is “record on the device, and report on the host”, simpleperf not only collects samples in profiling data, but also collects needed symbols, device info and recording time.

  2. It delivers new features for recording. a. When recording dwarf based call graph, simpleperf unwinds the stack before writing a sample to file. This is to save storage space on the device. b. Support tracing both on CPU time and off CPU time with --trace-offcpu option. c. Support recording callgraphs of JITed and interpreted Java code on Android >= P.

  3. It relates closely to the Android platform. a. Is aware of Android environment, like using system properties to enable profiling, using run-as to profile in application's context. b. Supports reading symbols and debug information from the .gnu_debugdata section, because system libraries are built with .gnu_debugdata section starting from Android O. c. Supports profiling shared libraries embedded in apk files. d. It uses the standard Android stack unwinder, so its results are consistent with all other Android tools.

  4. It builds executables and shared libraries for different usages. a. Builds static executables on the device. Since static executables don't rely on any library, simpleperf executables can be pushed on any Android device and used to record profiling data. b. Builds executables on different hosts: Linux, Mac and Windows. These executables can be used to report on hosts. c. Builds report shared libraries on different hosts. The report library is used by different Python scripts to parse profiling data.

Detailed documentation for the simpleperf executable is here.

Python scripts are split into three parts according to their functions:

  1. Scripts used for recording, like app_profiler.py, run_simpleperf_without_usb_connection.py.

  2. Scripts used for reporting, like report.py, report_html.py, inferno.

  3. Scripts used for parsing profiling data, like simpleperf_report_lib.py.

Detailed documentation for the Python scripts is here.

Tools in simpleperf

The simpleperf executables and Python scripts are located in simpleperf/ in ndk releases, and in system/extras/simpleperf/scripts/ in AOSP. Their functions are listed below.

bin/: contains executables and shared libraries.

bin/android/${arch}/simpleperf: static simpleperf executables used on the device.

bin/${host}/${arch}/simpleperf: simpleperf executables used on the host, only supports reporting.

bin/${host}/${arch}/libsimpleperf_report.${so/dylib/dll}: report shared libraries used on the host.

app_profiler.py: recording profiling data.

run_simpleperf_without_usb_connection.py: recording profiling data while the USB cable isn't connected.

binary_cache_builder.py: building binary cache for profiling data.

report.py: reporting in stdio interface.

report_html.py: reporting in html interface.

inferno.sh (or inferno.bat on Windows): generating flamegraph in html interface.

inferno/: implementation of inferno. Used by inferno.sh.

pprof_proto_generator.py: converting profiling data to the format used by pprof.

report_sample.py: converting profiling data to the format used by FlameGraph.

simpleperf_report_lib.py: library for parsing profiling data.

Android application profiling

This section shows how to profile an Android application. Some examples are Here.

Profiling an Android application involves three steps:

  1. Prepare an Android application.
  2. Record profiling data.
  3. Report profiling data.

Prepare an Android application

Based on the profiling situation, we may need to customize the build script to generate an apk file specifically for profiling. Below are some suggestions.

  1. If you want to profile a debug build of an application:

For the debug build type, Android studio sets android::debuggable=“true” in AndroidManifest.xml, enables JNI checks and may not optimize C/C++ code. It can be profiled by simpleperf without any change.

  1. If you want to profile a release build of an application:

For the release build type, Android studio sets android::debuggable=“false” in AndroidManifest.xml, disables JNI checks and optimizes C/C++ code. However, security restrictions mean that only apps with android::debuggable set to true can be profiled. So simpleperf can only profile a release build under these two circumstances: If you are on a rooted device, you can profile any app.

If you are on Android >= O, we can use wrap.sh to profile a release build: Step 1: Add android::debuggable=“true” in AndroidManifest.xml to enable profiling.

<manifest ...>
    <application android::debuggable="true" ...>

Step 2: Add wrap.sh in lib/arch directories. wrap.sh runs the app without passing any debug flags to ART, so the app runs as a release app. wrap.sh can be done by adding the script below in app/build.gradle.

android {
    buildTypes {
        release {
            sourceSets {
                release {
                    resources {
                        srcDir {
                            "wrap_sh_lib_dir"
                        }
                    }
                }
            }
        }
    }
}

task createWrapShLibDir
    for (String abi : ["armeabi", "armeabi-v7a", "arm64-v8a", "x86", "x86_64"]) {
        def dir = new File("app/wrap_sh_lib_dir/lib/" + abi)
        dir.mkdirs()
        def wrapFile = new File(dir, "wrap.sh")
        wrapFile.withWriter { writer ->
            writer.write('#!/system/bin/sh\n\$@\n')
        }
    }
}
  1. If you want to profile C/C++ code:

Android studio strips symbol table and debug info of native libraries in the apk. So the profiling results may contain unknown symbols or broken callgraphs. To fix this, we can pass app_profiler.py a directory containing unstripped native libraries via the -lib option. Usually the directory can be the path of your Android Studio project.

  1. If you want to profile Java code:

On Android >= P, simpleperf supports profiling Java code, no matter whether it is executed by the interpreter, or JITed, or compiled into native instructions. So you don't need to do anything.

On Android O, simpleperf supports profiling Java code which is compiled into native instructions, and it also needs wrap.sh to use the compiled Java code. To compile Java code, we can pass app_profiler.py the --compile_java_code option.

On Android N, simpleperf supports profiling Java code that is compiled into native instructions. To compile java code, we can pass app_profiler.py the --compile_java_code option.

On Android <= M, simpleperf doesn't support profiling Java code.

Below I use application SimpleperfExampleWithNative. It builds an app-profiling.apk for profiling.

$ git clone https://android.googlesource.com/platform/system/extras
$ cd extras/simpleperf/demo
# Open SimpleperfExamplesWithNative project with Android studio, and build this project
# successfully, otherwise the `./gradlew` command below will fail.
$ cd SimpleperfExampleWithNative

# On windows, use "gradlew" instead.
$ ./gradlew clean assemble
$ adb install -r app/build/outputs/apk/profiling/app-profiling.apk

Record and report profiling data

We can use app-profiler.py to profile Android applications.

# Cd to the directory of simpleperf scripts. Record perf.data.
# -p option selects the profiled app using its package name.
# --compile_java_code option compiles Java code into native instructions, which isn't needed on
# Android >= P.
# -a option selects the Activity to profile.
# -lib option gives the directory to find debug native libraries.
$ python app_profiler.py -p com.example.simpleperf.simpleperfexamplewithnative --compile_java_code \
    -a .MixActivity -lib path_of_SimpleperfExampleWithNative

This will collect profiling data in perf.data in the current directory, and related native binaries in binary_cache/.

Normally we need to use the app when profiling, otherwise we may record no samples. But in this case, the MixActivity starts a busy thread. So we don't need to use the app while profiling.

# Report perf.data in stdio interface.
$ python report.py
Cmdline: /data/data/com.example.simpleperf.simpleperfexamplewithnative/simpleperf record ...
Arch: arm64
Event: task-clock:u (type 1, config 1)
Samples: 10023
Event count: 10023000000

Overhead  Command     Pid   Tid   Shared Object              Symbol
27.04%    BusyThread  5703  5729  /system/lib64/libart.so    art::JniMethodStart(art::Thread*)
25.87%    BusyThread  5703  5729  /system/lib64/libc.so      long StrToI<long, ...
...

report.py reports profiling data in stdio interface. If there are a lot of unknown symbols in the report, check here.

# Report perf.data in html interface.
$ python report_html.py

# Add source code and disassembly. Change the path of source_dirs if it not correct.
$ python report_html.py --add_source_code --source_dirs path_of_SimpleperfExampleWithNative \
      --add_disassembly

report_html.py generates report in report.html, and pops up a browser tab to show it.

Record and report call graph

We can record and report call graphs as below.

# Record dwarf based call graphs: add "-g" in the -r option.
$ python app_profiler.py -p com.example.simpleperf.simpleperfexamplewithnative \
        -r "-e task-clock:u -f 1000 --duration 10 -g" -lib path_of_SimpleperfExampleWithNative

# Record stack frame based call graphs: add "--call-graph fp" in the -r option.
$ python app_profiler.py -p com.example.simpleperf.simpleperfexamplewithnative \
        -r "-e task-clock:u -f 1000 --duration 10 --call-graph fp" \
        -lib path_of_SimpleperfExampleWithNative

# Report call graphs in stdio interface.
$ python report.py -g

# Report call graphs in python Tk interface.
$ python report.py -g --gui

# Report call graphs in html interface.
$ python report_html.py

# Report call graphs in flame graphs.
# On Windows, use inferno.bat instead of ./inferno.sh.
$ ./inferno.sh -sc

Report in html interface

We can use report_html.py to show profiling results in a web browser. report_html.py integrates chart statistics, sample table, flame graphs, source code annotation and disassembly annotation. It is the recommended way to show reports.

$ python report_html.py

Show flame graph

To show flame graphs, we need to first record call graphs. Flame graphs are shown by report_html.py in the “Flamegraph” tab. We can also use inferno to show flame graphs directly.

# On Windows, use inferno.bat instead of ./inferno.sh.
$ ./inferno.sh -sc

We can also build flame graphs using https://github.com/brendangregg/FlameGraph. Please make sure you have perl installed.

$ git clone https://github.com/brendangregg/FlameGraph.git
$ python report_sample.py --symfs binary_cache >out.perf
$ FlameGraph/stackcollapse-perf.pl out.perf >out.folded
$ FlameGraph/flamegraph.pl out.folded >a.svg

Record both on CPU time and off CPU time

We can record both on CPU time and off CPU time.

First check if trace-offcpu feature is supported on the device.

$ python run_simpleperf_on_device.py list --show-features
dwarf-based-call-graph
trace-offcpu

If trace-offcpu is supported, it will be shown in the feature list. Then we can try it.

$ python app_profiler.py -p com.example.simpleperf.simpleperfexamplewithnative -a .SleepActivity \
    -r "-g -e task-clock:u -f 1000 --duration 10 --trace-offcpu" \
    -lib path_of_SimpleperfExampleWithNative
$ python report_html.py --add_disassembly --add_source_code \
    --source_dirs path_of_SimpleperfExampleWithNative

Profile from launch

We can profile from launch of an application.

# Start simpleperf recording, then start the Activity to profile.
$ python app_profiler.py -p com.example.simpleperf.simpleperfexamplewithnative -a .MainActivity

# We can also start the Activity on the device manually.
# 1. Make sure the application isn't running or one of the recent apps.
# 2. Start simpleperf recording.
$ python app_profiler.py -p com.example.simpleperf.simpleperfexamplewithnative
# 3. Start the app manually on the device.

Parse profiling data manually

We can also write python scripts to parse profiling data manually, by using simpleperf_report_lib.py. Examples are report_sample.py, report_html.py.

Android platform profiling

Here are some tips for Android platform developers, who build and flash system images on rooted devices:

  1. After running adb root, simpleperf can be used to profile any process or system wide.
  2. It is recommended to use the latest simpleperf available in AOSP master, if you are not working on the current master branch. Scripts are in system/extras/simpleperf/scripts, binaries are in system/extras/simpleperf/scripts/bin/android.
  3. It is recommended to use app_profiler.py for recording, and report_html.py for reporting. Below is an example.
# Record surfaceflinger process for 10 seconds with dwarf based call graph. More examples are in
# scripts reference in the doc.
$ python app_profiler.py -np surfaceflinger -r "-g --duration 10"

# Generate html report.
$ python report_html.py
  1. Since Android >= O has symbols for system libraries on device, we don't need to use unstripped binaries in $ANDROID_PRODUCT_OUT/symbols to report call graphs. However, they are needed to add source code and disassembly (with line numbers) in the report. Below is an example.
# Doing recording with app_profiler.py or simpleperf on device, and generates perf.data on host.
$ python app_profiler.py -np surfaceflinger -r "--call-graph fp --duration 10"

# Collect unstripped binaries from $ANDROID_PRODUCT_OUT/symbols to binary_cache/.
$ python binary_cache_builder.py -lib $ANDROID_PRODUCT_OUT/symbols

# Report source code and disassembly. Disassembling all binaries is slow, so it's better to add
# --binary_filter option to only disassemble selected binaries.
$ python report_html.py --add_source_code --source_dirs $ANDROID_BUILD_TOP --add_disassembly \
  --binary_filter surfaceflinger.so

Executable commands reference

How simpleperf works

Modern CPUs have a hardware component called the performance monitoring unit (PMU). The PMU has several hardware counters, counting events like how many cpu cycles have happened, how many instructions have executed, or how many cache misses have happened.

The Linux kernel wraps these hardware counters into hardware perf events. In addition, the Linux kernel also provides hardware independent software events and tracepoint events. The Linux kernel exposes all events to userspace via the perf_event_open system call, which is used by simpleperf.

Simpleperf has three main commands: stat, record and report.

The stat command gives a summary of how many events have happened in the profiled processes in a time period. Here’s how it works:

  1. Given user options, simpleperf enables profiling by making a system call to the kernel.
  2. The kernel enables counters while the profiled processes are running.
  3. After profiling, simpleperf reads counters from the kernel, and reports a counter summary.

The record command records samples of the profiled processes in a time period. Here’s how it works:

  1. Given user options, simpleperf enables profiling by making a system call to the kernel.
  2. Simpleperf creates mapped buffers between simpleperf and the kernel.
  3. The kernel enables counters while the profiled processes are running.
  4. Each time a given number of events happen, the kernel dumps a sample to the mapped buffers.
  5. Simpleperf reads samples from the mapped buffers and stores profiling data in a file called perf.data.

The report command reads perf.data and any shared libraries used by the profiled processes, and outputs a report showing where the time was spent.

Commands

Simpleperf supports several commands, listed below:

The debug-unwind command: debug/test dwarf based offline unwinding, used for debugging simpleperf.
The dump command: dumps content in perf.data, used for debugging simpleperf.
The help command: prints help information for other commands.
The kmem command: collects kernel memory allocation information (will be replaced by Python scripts).
The list command: lists all event types supported on the Android device.
The record command: profiles processes and stores profiling data in perf.data.
The report command: reports profiling data in perf.data.
The report-sample command: reports each sample in perf.data, used for supporting integration of
                           simpleperf in Android Studio.
The stat command: profiles processes and prints counter summary.

Each command supports different options, which can be seen through help message.

# List all commands.
$ simpleperf --help

# Print help message for record command.
$ simpleperf record --help

Below describes the most frequently used commands, which are list, stat, record and report.

The list command

The list command lists all events available on the device. Different devices may support different events because they have different hardware and kernels.

$ simpleperf list
List of hw-cache events:
  branch-loads
  ...
List of hardware events:
  cpu-cycles
  instructions
  ...
List of software events:
  cpu-clock
  task-clock
  ...

On ARM/ARM64, the list command also shows a list of raw events, they are the events supported by the ARM PMU on the device. The kernel has wrapped part of them into hardware events and hw-cache events. For example, raw-cpu-cycles is wrapped into cpu-cycles, raw-instruction-retired is wrapped into instructions. The raw events are provided in case we want to use some events supported on the device, but unfortunately not wrapped by the kernel.

The stat command

The stat command is used to get event counter values of the profiled processes. By passing options, we can select which events to use, which processes/threads to monitor, how long to monitor and the print interval.

# Stat using default events (cpu-cycles,instructions,...), and monitor process 7394 for 10 seconds.
$ simpleperf stat -p 7394 --duration 10
Performance counter statistics:

 1,320,496,145  cpu-cycles         # 0.131736 GHz                     (100%)
   510,426,028  instructions       # 2.587047 cycles per instruction  (100%)
     4,692,338  branch-misses      # 468.118 K/sec                    (100%)
886.008130(ms)  task-clock         # 0.088390 cpus used               (100%)
           753  context-switches   # 75.121 /sec                      (100%)
           870  page-faults        # 86.793 /sec                      (100%)

Total test time: 10.023829 seconds.

Select events to stat

We can select which events to use via -e.

# Stat event cpu-cycles.
$ simpleperf stat -e cpu-cycles -p 11904 --duration 10

# Stat event cache-references and cache-misses.
$ simpleperf stat -e cache-references,cache-misses -p 11904 --duration 10

When running the stat command, if the number of hardware events is larger than the number of hardware counters available in the PMU, the kernel shares hardware counters between events, so each event is only monitored for part of the total time. In the example below, there is a percentage at the end of each row, showing the percentage of the total time that each event was actually monitored.

# Stat using event cache-references, cache-references:u,....
$ simpleperf stat -p 7394 -e cache-references,cache-references:u,cache-references:k \
      -e cache-misses,cache-misses:u,cache-misses:k,instructions --duration 1
Performance counter statistics:

4,331,018  cache-references     # 4.861 M/sec    (87%)
3,064,089  cache-references:u   # 3.439 M/sec    (87%)
1,364,959  cache-references:k   # 1.532 M/sec    (87%)
   91,721  cache-misses         # 102.918 K/sec  (87%)
   45,735  cache-misses:u       # 51.327 K/sec   (87%)
   38,447  cache-misses:k       # 43.131 K/sec   (87%)
9,688,515  instructions         # 10.561 M/sec   (89%)

Total test time: 1.026802 seconds.

In the example above, each event is monitored about 87% of the total time. But there is no guarantee that any pair of events are always monitored at the same time. If we want to have some events monitored at the same time, we can use --group.

# Stat using event cache-references, cache-references:u,....
$ simpleperf stat -p 7964 --group cache-references,cache-misses \
      --group cache-references:u,cache-misses:u --group cache-references:k,cache-misses:k \
      -e instructions --duration 1
Performance counter statistics:

3,638,900  cache-references     # 4.786 M/sec          (74%)
   65,171  cache-misses         # 1.790953% miss rate  (74%)
2,390,433  cache-references:u   # 3.153 M/sec          (74%)
   32,280  cache-misses:u       # 1.350383% miss rate  (74%)
  879,035  cache-references:k   # 1.251 M/sec          (68%)
   30,303  cache-misses:k       # 3.447303% miss rate  (68%)
8,921,161  instructions         # 10.070 M/sec         (86%)

Total test time: 1.029843 seconds.

Select target to stat

We can select which processes or threads to monitor via -p or -t. Monitoring a process is the same as monitoring all threads in the process. Simpleperf can also fork a child process to run the new command and then monitor the child process.

# Stat process 11904 and 11905.
$ simpleperf stat -p 11904,11905 --duration 10

# Stat thread 11904 and 11905.
$ simpleperf stat -t 11904,11905 --duration 10

# Start a child process running `ls`, and stat it.
$ simpleperf stat ls

# Stat the process of an Android application. This only works for debuggable apps on non-rooted
# devices.
$ simpleperf stat --app com.example.simpleperf.simpleperfexamplewithnative

# Stat system wide using -a.
$ simpleperf stat -a --duration 10

Decide how long to stat

When monitoring existing threads, we can use --duration to decide how long to monitor. When monitoring a child process running a new command, simpleperf monitors until the child process ends. In this case, we can use Ctrl-C to stop monitoring at any time.

# Stat process 11904 for 10 seconds.
$ simpleperf stat -p 11904 --duration 10

# Stat until the child process running `ls` finishes.
$ simpleperf stat ls

# Stop monitoring using Ctrl-C.
$ simpleperf stat -p 11904 --duration 10
^C

If you want to write a script to control how long to monitor, you can send one of SIGINT, SIGTERM, SIGHUP signals to simpleperf to stop monitoring.

Decide the print interval

When monitoring perf counters, we can also use --interval to decide the print interval.

# Print stat for process 11904 every 300ms.
$ simpleperf stat -p 11904 --duration 10 --interval 300

# Print system wide stat at interval of 300ms for 10 seconds. Note that system wide profiling needs
# root privilege.
$ su 0 simpleperf stat -a --duration 10 --interval 300

Display counters in systrace

Simpleperf can also work with systrace to dump counters in the collected trace. Below is an example to do a system wide stat.

# Capture instructions (kernel only) and cache misses with interval of 300 milliseconds for 15
# seconds.
$ su 0 simpleperf stat -e instructions:k,cache-misses -a --interval 300 --duration 15
# On host launch systrace to collect trace for 10 seconds.
(HOST)$ external/chromium-trace/systrace.py --time=10 -o new.html sched gfx view
# Open the collected new.html in browser and perf counters will be shown up.

The record command

The record command is used to dump samples of the profiled processes. Each sample can contain information like the time at which the sample was generated, the number of events since last sample, the program counter of a thread, the call chain of a thread.

By passing options, we can select which events to use, which processes/threads to monitor, what frequency to dump samples, how long to monitor, and where to store samples.

# Record on process 7394 for 10 seconds, using default event (cpu-cycles), using default sample
# frequency (4000 samples per second), writing records to perf.data.
$ simpleperf record -p 7394 --duration 10
simpleperf I cmd_record.cpp:316] Samples recorded: 21430. Samples lost: 0.

Select events to record

By default, the cpu-cycles event is used to evaluate consumed cpu cycles. But we can also use other events via -e.

# Record using event instructions.
$ simpleperf record -e instructions -p 11904 --duration 10

# Record using task-clock, which shows the passed CPU time in nanoseconds.
$ simpleperf record -e task-clock -p 11904 --duration 10

Select target to record

The way to select target in record command is similar to that in the stat command.

# Record process 11904 and 11905.
$ simpleperf record -p 11904,11905 --duration 10

# Record thread 11904 and 11905.
$ simpleperf record -t 11904,11905 --duration 10

# Record a child process running `ls`.
$ simpleperf record ls

# Record the process of an Android application. This only works for debuggable apps on non-rooted
# devices.
$ simpleperf record --app com.example.simpleperf.simpleperfexamplewithnative

# Record system wide.
$ simpleperf record -a --duration 10

Set the frequency to record

We can set the frequency to dump records via -f or -c. For example, -f 4000 means dumping approximately 4000 records every second when the monitored thread runs. If a monitored thread runs 0.2s in one second (it can be preempted or blocked in other times), simpleperf dumps about 4000 * 0.2 / 1.0 = 800 records every second. Another way is using -c. For example, -c 10000 means dumping one record whenever 10000 events happen.

# Record with sample frequency 1000: sample 1000 times every second running.
$ simpleperf record -f 1000 -p 11904,11905 --duration 10

# Record with sample period 100000: sample 1 time every 100000 events.
$ simpleperf record -c 100000 -t 11904,11905 --duration 10

To avoid taking too much time generating samples, kernel >= 3.10 sets the max percent of cpu time used for generating samples (default is 25%), and decreases the max allowed sample frequency when hitting that limit. Simpleperf uses --cpu-percent option to adjust it, but it needs either root privilege or to be on Android >= Q.

# Record with sample frequency 10000, with max allowed cpu percent to be 50%.
$ simpleperf record -f 1000 -p 11904,11905 --duration 10 --cpu-percent 50

Decide how long to record

The way to decide how long to monitor in record command is similar to that in the stat command.

# Record process 11904 for 10 seconds.
$ simpleperf record -p 11904 --duration 10

# Record until the child process running `ls` finishes.
$ simpleperf record ls

# Stop monitoring using Ctrl-C.
$ simpleperf record -p 11904 --duration 10
^C

If you want to write a script to control how long to monitor, you can send one of SIGINT, SIGTERM, SIGHUP signals to simpleperf to stop monitoring.

Set the path to store profiling data

By default, simpleperf stores profiling data in perf.data in the current directory. But the path can be changed using -o.

# Write records to data/perf2.data.
$ simpleperf record -p 11904 -o data/perf2.data --duration 10

Record call graphs

A call graph is a tree showing function call relations. Below is an example.

main() {
    FunctionOne();
    FunctionTwo();
}
FunctionOne() {
    FunctionTwo();
    FunctionThree();
}
a call graph:
    main-> FunctionOne
       |    |
       |    |-> FunctionTwo
       |    |-> FunctionThree
       |
       |-> FunctionTwo

A call graph shows how a function calls other functions, and a reversed call graph shows how a function is called by other functions. To show a call graph, we need to first record it, then report it.

There are two ways to record a call graph, one is recording a dwarf based call graph, the other is recording a stack frame based call graph. Recording dwarf based call graphs needs support of debug information in native binaries. While recording stack frame based call graphs needs support of stack frame registers.

# Record a dwarf based call graph
$ simpleperf record -p 11904 -g --duration 10

# Record a stack frame based call graph
$ simpleperf record -p 11904 --call-graph fp --duration 10

Here are some suggestions about recording call graphs.

Record both on CPU time and off CPU time

Simpleperf is a CPU profiler, it generates samples for a thread only when it is running on a CPU. However, sometimes we want to figure out where the time of a thread is spent, whether it is running on a CPU, or staying in the kernel's ready queue, or waiting for something like I/O events.

To support this, the record command uses --trace-offcpu to trace both on CPU time and off CPU time. When --trace-offcpu is used, simpleperf generates a sample when a running thread is scheduled out, so we know the callstack of a thread when it is scheduled out. And when reporting a perf.data generated with --trace-offcpu, we use time to the next sample (instead of event counts from the previous sample) as the weight of the current sample. As a result, we can get a call graph based on timestamps, including both on CPU time and off CPU time.

trace-offcpu is implemented using sched:sched_switch tracepoint event, which may not be supported on old kernels. But it is guaranteed to be supported on devices >= Android O MR1. We can check whether trace-offcpu is supported as below.

$ simpleperf list --show-features
dwarf-based-call-graph
trace-offcpu

If trace-offcpu is supported, it will be shown in the feature list. Then we can try it.

# Record with --trace-offcpu.
$ simpleperf record -g -p 11904 --duration 10 --trace-offcpu

# Record with --trace-offcpu using app_profiler.py.
$ python app_profiler.py -p com.example.simpleperf.simpleperfexamplewithnative -a .SleepActivity \
    -r "-g -e task-clock:u -f 1000 --duration 10 --trace-offcpu"

Below is an example comparing the profiling result with / without --trace-offcpu. First we record without --trace-offcpu.

$ python app_profiler.py -p com.example.simpleperf.simpleperfexamplewithnative -a .SleepActivity

$ python report_html.py --add_disassembly --add_source_code --source_dirs ../demo

The result is here. In the result, all time is taken by RunFunction(), and sleep time is ignored. But if we add --trace-offcpu, the result changes.

$ python app_profiler.py -p com.example.simpleperf.simpleperfexamplewithnative -a .SleepActivity \
    -r "-g -e task-clock:u --trace-offcpu -f 1000 --duration 10"

$ python report_html.py --add_disassembly --add_source_code --source_dirs ../demo

The result is here. In the result, half of the time is taken by RunFunction(), and the other half is taken by SleepFunction(). So it traces both on CPU time and off CPU time.

The report command

The report command is used to report profiling data generated by the record command. The report contains a table of sample entries. Each sample entry is a row in the report. The report command groups samples belong to the same process, thread, library, function in the same sample entry. Then sort the sample entries based on the event count a sample entry has.

By passing options, we can decide how to filter out uninteresting samples, how to group samples into sample entries, and where to find profiling data and binaries.

Below is an example. Records are grouped into 4 sample entries, each entry is a row. There are several columns, each column shows piece of information belonging to a sample entry. The first column is Overhead, which shows the percentage of events inside the current sample entry in total events. As the perf event is cpu-cycles, the overhead is the percentage of CPU cycles used in each function.

# Reports perf.data, using only records sampled in libsudo-game-jni.so, grouping records using
# thread name(comm), process id(pid), thread id(tid), function name(symbol), and showing sample
# count for each row.
$ simpleperf report --dsos /data/app/com.example.sudogame-2/lib/arm64/libsudo-game-jni.so \
      --sort comm,pid,tid,symbol -n
Cmdline: /data/data/com.example.sudogame/simpleperf record -p 7394 --duration 10
Arch: arm64
Event: cpu-cycles (type 0, config 0)
Samples: 28235
Event count: 546356211

Overhead  Sample  Command    Pid   Tid   Symbol
59.25%    16680   sudogame  7394  7394  checkValid(Board const&, int, int)
20.42%    5620    sudogame  7394  7394  canFindSolution_r(Board&, int, int)
13.82%    4088    sudogame  7394  7394  randomBlock_r(Board&, int, int, int, int, int)
6.24%     1756    sudogame  7394  7394  @plt

Set the path to read profiling data

By default, the report command reads profiling data from perf.data in the current directory. But the path can be changed using -i.

$ simpleperf report -i data/perf2.data

Set the path to find binaries

To report function symbols, simpleperf needs to read executable binaries used by the monitored processes to get symbol table and debug information. By default, the paths are the executable binaries used by monitored processes while recording. However, these binaries may not exist when reporting or not contain symbol table and debug information. So we can use --symfs to redirect the paths.

# In this case, when simpleperf wants to read executable binary /A/b, it reads file in /A/b.
$ simpleperf report

# In this case, when simpleperf wants to read executable binary /A/b, it prefers file in
# /debug_dir/A/b to file in /A/b.
$ simpleperf report --symfs /debug_dir

# Read symbols for system libraries built locally. Note that this is not needed since Android O,
# which ships symbols for system libraries on device.
$ simpleperf report --symfs $ANDROID_PRODUCT_OUT/symbols

Filter samples

When reporting, it happens that not all records are of interest. The report command supports four filters to select samples of interest.

# Report records in threads having name sudogame.
$ simpleperf report --comms sudogame

# Report records in process 7394 or 7395
$ simpleperf report --pids 7394,7395

# Report records in thread 7394 or 7395.
$ simpleperf report --tids 7394,7395

# Report records in libsudo-game-jni.so.
$ simpleperf report --dsos /data/app/com.example.sudogame-2/lib/arm64/libsudo-game-jni.so

Group samples into sample entries

The report command uses --sort to decide how to group sample entries.

# Group records based on their process id: records having the same process id are in the same
# sample entry.
$ simpleperf report --sort pid

# Group records based on their thread id and thread comm: records having the same thread id and
# thread name are in the same sample entry.
$ simpleperf report --sort tid,comm

# Group records based on their binary and function: records in the same binary and function are in
# the same sample entry.
$ simpleperf report --sort dso,symbol

# Default option: --sort comm,pid,tid,dso,symbol. Group records in the same thread, and belong to
# the same function in the same binary.
$ simpleperf report

Report call graphs

To report a call graph, please make sure the profiling data is recorded with call graphs, as here.

$ simpleperf report -g

Scripts reference

app_profiler.py

app_profiler.py is used to record profiling data for Android applications and native executables.

# Record an Android application.
$ python app_profiler.py -p com.example.simpleperf.simpleperfexamplewithnative

# Record an Android application with Java code compiled into native instructions.
$ python app_profiler.py -p com.example.simpleperf.simpleperfexamplewithnative --compile_java_code

# Record the launch of an Activity of an Android application.
$ python app_profiler.py -p com.example.simpleperf.simpleperfexamplewithnative -a .SleepActivity

# Record a native process.
$ python app_profiler.py -np surfaceflinger

# Record a native process given its pid.
$ python app_profiler.py --pid 11324

# Record a command.
$ python app_profiler.py -cmd \
    "dex2oat --dex-file=/data/local/tmp/app-profiling.apk --oat-file=/data/local/tmp/a.oat"

# Record an Android application, and use -r to send custom options to the record command.
$ python app_profiler.py -p com.example.simpleperf.simpleperfexamplewithnative \
    -r "-e cpu-clock -g --duration 30"

# Record both on CPU time and off CPU time.
$ python app_profiler.py -p com.example.simpleperf.simpleperfexamplewithnative \
    -r "-e task-clock -g -f 1000 --duration 10 --trace-offcpu"

# Save profiling data in a custom file (like perf_custom.data) instead of perf.data.
$ python app_profiler.py -p com.example.simpleperf.simpleperfexamplewithnative -o perf_custom.data

Profile from launch of an application

Sometimes we want to profile the launch-time of an application. To support this, we added --app in the record command. The --app option sets the package name of the Android application to profile. If the app is not already running, the record command will poll for the app process in a loop with an interval of 1ms. So to profile from launch of an application, we can first start the record command with --app, then start the app. Below is an example.

$ python run_simpleperf_on_device.py record
    --app com.example.simpleperf.simpleperfexamplewithnative \
    -g --duration 1 -o /data/local/tmp/perf.data
# Start the app manually or using the `am` command.

To make it convenient to use, app_profiler.py supports using the -a option to start an Activity after recording has started.

$ python app_profiler.py -p com.example.simpleperf.simpleperfexamplewithnative -a .MainActivity

run_simpleperf_without_usb_connection.py

run_simpleperf_without_usb_connection.py records profiling data while the USB cable isn't connected. Below is an example.

$ python run_simpleperf_without_usb_connection.py start \
    -p com.example.simpleperf.simpleperfexamplewithnative
# After the command finishes successfully, unplug the USB cable, run the
# SimpleperfExampleWithNative app. After a few seconds, plug in the USB cable.
$ python run_simpleperf_without_usb_connection.py stop
# It may take a while to stop recording. After that, the profiling data is collected in perf.data
# on host.

binary_cache_builder.py

The binary_cache directory is a directory holding binaries needed by a profiling data file. The binaries are expected to be unstripped, having debug information and symbol tables. The binary_cache directory is used by report scripts to read symbols of binaries. It is also used by report_html.py to generate annotated source code and disassembly.

By default, app_profiler.py builds the binary_cache directory after recording. But we can also build binary_cache for existing profiling data files using binary_cache_builder.py. It is useful when you record profiling data using simpleperf record directly, to do system wide profiling or record without the USB cable connected.

binary_cache_builder.py can either pull binaries from an Android device, or find binaries in directories on the host (via -lib).

# Generate binary_cache for perf.data, by pulling binaries from the device.
$ python binary_cache_builder.py

# Generate binary_cache, by pulling binaries from the device and finding binaries in
# SimpleperfExampleWithNative.
$ python binary_cache_builder.py -lib path_of_SimpleperfExampleWithNative

run_simpleperf_on_device.py

This script pushes the simpleperf executable on the device, and run a simpleperf command on the device. It is more convenient than running adb commands manually.

report.py

report.py is a wrapper of the report command on the host. It accepts all options of the report command.

# Report call graph
$ python report.py -g

# Report call graph in a GUI window implemented by Python Tk.
$ python report.py -g --gui

report_html.py

report_html.py generates report.html based on the profiling data. Then the report.html can show the profiling result without depending on other files. So it can be shown in local browsers or passed to other machines. Depending on which command-line options are used, the content of the report.html can include: chart statistics, sample table, flame graphs, annotated source code for each function, annotated disassembly for each function.

# Generate chart statistics, sample table and flame graphs, based on perf.data.
$ python report_html.py

# Add source code.
$ python report_html.py --add_source_code --source_dirs path_of_SimpleperfExampleWithNative

# Add disassembly.
$ python report_html.py --add_disassembly

# Adding disassembly for all binaries can cost a lot of time. So we can choose to only add
# disassembly for selected binaries.
$ python report_html.py --add_disassembly --binary_filter libgame.so

# report_html.py accepts more than one recording data file.
$ python report_html.py -i perf1.data perf2.data

Below is an example of generating html profiling results for SimpleperfExampleWithNative.

$ python app_profiler.py -p com.example.simpleperf.simpleperfexamplewithnative
$ python report_html.py --add_source_code --source_dirs path_of_SimpleperfExampleWithNative \
    --add_disassembly

After opening the generated report.html in a browser, there are several tabs:

The first tab is “Chart Statistics”. You can click the pie chart to show the time consumed by each process, thread, library and function.

The second tab is “Sample Table”. It shows the time taken by each function. By clicking one row in the table, we can jump to a new tab called “Function”.

The third tab is “Flamegraph”. It shows the flame graphs generated by inferno.

The fourth tab is “Function”. It only appears when users click a row in the “Sample Table” tab. It shows information of a function, including:

  1. A flame graph showing functions called by that function.
  2. A flame graph showing functions calling that function.
  3. Annotated source code of that function. It only appears when there are source code files for that function.
  4. Annotated disassembly of that function. It only appears when there are binaries containing that function.

inferno

inferno is a tool used to generate flame graph in a html file.

# Generate flame graph based on perf.data.
# On Windows, use inferno.bat instead of ./inferno.sh.
$ ./inferno.sh -sc --record_file perf.data

# Record a native program and generate flame graph.
$ ./inferno.sh -np surfaceflinger

pprof_proto_generator.py

It converts a profiling data file into pprof.proto, a format used by pprof.

# Convert perf.data in the current directory to pprof.proto format.
$ python pprof_proto_generator.py
$ pprof -pdf pprof.profile

report_sample.py

It converts a profiling data file into a format used by FlameGraph.

# Convert perf.data in the current directory to a format used by FlameGraph.
$ python report_sample.py --symfs binary_cache >out.perf
$ git clone https://github.com/brendangregg/FlameGraph.git
$ FlameGraph/stackcollapse-perf.pl out.perf >out.folded
$ FlameGraph/flamegraph.pl out.folded >a.svg

simpleperf_report_lib.py

simpleperf_report_lib.py is a Python library used to parse profiling data files generated by the record command. Internally, it uses libsimpleperf_report.so to do the work. Generally, for each profiling data file, we create an instance of ReportLib, pass it the file path (via SetRecordFile). Then we can read all samples through GetNextSample(). For each sample, we can read its event info (via GetEventOfCurrentSample), symbol info (via GetSymbolOfCurrentSample) and call chain info (via GetCallChainOfCurrentSample). We can also get some global information, like record options (via GetRecordCmd), the arch of the device (via GetArch) and meta strings (via MetaInfo).

Examples of using simpleperf_report_lib.py are in report_sample.py, report_html.py, pprof_proto_generator.py and inferno/inferno.py.

Answers to common issues

Why we suggest profiling on Android >= N devices?

1. Running on a device reflects a real running situation, so we suggest
profiling on real devices instead of emulators.
2. To profile Java code, we need ART running in oat mode, which is only
available >= L for rooted devices, and >= N for non-rooted devices.
3. Old Android versions are likely to be shipped with old kernels (< 3.18),
which may not support profiling features like recording dwarf based call graphs.
4. Old Android versions are likely to be shipped with Arm32 chips. In Arm32
mode, recording stack frame based call graphs doesn't work well.

Suggestions about recording call graphs

Below is our experiences of dwarf based call graphs and stack frame based call graphs.

dwarf based call graphs:

  1. Need support of debug information in binaries.
  2. Behave normally well on both ARM and ARM64, for both fully compiled Java code and C++ code.
  3. Can only unwind 64K stack for each sample. So usually can't show complete flame-graph. But probably is enough for users to identify hot places.
  4. Take more CPU time than stack frame based call graphs. So the sample frequency is suggested to be 1000 Hz. Thus at most 1000 samples per second.

stack frame based call graphs:

  1. Need support of stack frame registers.
  2. Don‘t work well on ARM. Because ARM is short of registers, and ARM and THUMB code have different stack frame registers. So the kernel can’t unwind user stack containing both ARM/THUMB code.
  3. Also don‘t work well on fully compiled Java code on ARM64. Because the ART compiler doesn’t reserve stack frame registers.
  4. Work well when profiling native programs on ARM64. One example is profiling surfacelinger. And usually shows complete flame-graph when it works well.
  5. Take less CPU time than dwarf based call graphs. So the sample frequency can be 4000 Hz or higher.

So if you need to profile code on ARM or profile fully compiled Java code, dwarf based call graphs may be better. If you need to profile C++ code on ARM64, stack frame based call graphs may be better. After all, you can always try dwarf based call graph first, because it always produces reasonable results when given unstripped binaries properly. If it doesn't work well enough, then try stack frame based call graphs instead.

Simpleperf needs to have unstripped native binaries on the device to generate good dwarf based call graphs. It can be supported in two ways:

  1. Use unstripped native binaries when building the apk, as here.
  2. Pass directory containing unstripped native libraries to app_profiler.py via -lib. And it will download the unstripped native libraries on the device.
$ python app_profiler.py -lib NATIVE_LIB_DIR

How to solve missing symbols in report?

The simpleperf record command collects symbols on device in perf.data. But if the native libraries you use on device are stripped, this will result in a lot of unknown symbols in the report. A solution is to build binary_cache on host.

# Collect binaries needed by perf.data in binary_cache/.
$ python binary_cache_builder.py -lib NATIVE_LIB_DIR,...

The NATIVE_LIB_DIRs passed in -lib option are the directories containing unstripped native libraries on host. After running it, the native libraries containing symbol tables are collected in binary_cache/ for use when reporting.

$ python report.py --symfs binary_cache

# report_html.py searches binary_cache/ automatically, so you don't need to
# pass it any argument.
$ python report_html.py

Bugs and contribution

Bugs and feature requests can be submitted at http://github.com/android-ndk/ndk/issues. Patches can be uploaded to android-review.googlesource.com as here, or sent to email addresses listed here.

If you want to compile simpleperf C++ source code, follow below steps:

  1. Download AOSP master branch as here.
  2. Build simpleperf.
$ . build/envsetup.sh
$ lunch aosp_arm64-userdebug
$ mmma system/extras/simpleperf -j30

If built successfully, out/target/product/generic_arm64/system/bin/simpleperf is for ARM64, and out/target/product/generic_arm64/system/bin/simpleperf32 is for ARM.