Scribd
Upload
4 views21 pages

Introduction

Simpleperf is a CPU profiler for Android that replaces the linux/tools/perf utility, utilizing Linux kernel support and hardware performance monitoring units (PMUs). It provides various commands for listing available events, monitoring performance, recording data, and reporting results, with detailed interactions between userspace, kernel, and hardware. The documentation and source code are available online, and it supports multiple options for profiling and data collection.

Uploaded by

Dedy UU
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
4 views21 pages

Introduction

Simpleperf is a CPU profiler for Android that replaces the linux/tools/perf utility, utilizing Linux kernel support and hardware performance monitoring units (PMUs). It provides various commands for listing available events, monitoring performance, recording data, and reporting results, with detailed interactions between userspace, kernel, and hardware. The documentation and source code are available online, and it supports multiple options for profiling and data collection.

Uploaded by

Dedy UU
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 21

Simpleperf Introduction

Yabin Cui
android-llvm-dev
Outline
● What is simpleperf
● How simpleperf works
● Simpleperf commands
What is simpleperf
● A replacement for linux/tools/perf in Android
● A cpu-profiler using linux kernel support and PMU (performance monitor unit)
hardware support
● Source code is in
https://android.googlesource.com/platform/system/extras/+/master/simpleperf/
● Doc is in
https://android.googlesource.com/platform/system/extras/+/master/simpleperf/doc/
● Prebuilt is release in
https://android.googlesource.com/platform/prebuilts/simpleperf/
How simpleperf works

Userspace Kernel Hardware

cpu pmu driver ARM PMU


simpleperf stat cmd

software pmu driver


simpleperf record cmd perf
event tracepoint pmu
driver driver
profile data file
device specific pmu
simpleperf report cmd drivers
and other report
methods
ARM PMU
● Described in ARM manual, D7 The Performance Monitors Extension
● PMU counter: Each cpu core has several PMU counters. Each counter is 32-bit, can monitor one
PMU event. When the monitored event happens, the counter value increases by one. When a
counter overflows, it can trigger an interrupt.
● PMU event: like CPU_CYCLES, BR_PRED (predictable branch), L1D_CACHE (Level 1 data cache
access). ARM lists common events and how to interpret them. And the events can be used together
to get indirect information, like cache miss rate = cache refill count / cache_access_count.
● The PMU events are growing in newer architectures.
How simpleperf works

Userspace Kernel Hardware

cpu pmu driver ARM PMU


simpleperf stat cmd

software pmu driver


simpleperf record cmd perf
event tracepoint pmu
driver driver
profile data file
device specific pmu
simpleperf report cmd drivers
and other report
methods
Kernel support
● perf event driver
○ a bridge between userspace and pmu drivers. It lives in kernel/events
○ maps pmu events to perf event types, described in include/uapi/linux/perf_event.h
○ provides a sysfs interface to show supported perf events, in /sys/bus/event_source
○ provides perf_event_open system call to monitor performance of selected threads

int perf_event_open(struct perf_event_attr *attr, pid_t pid, int cpu, int group_fd, unsigned
long flags)

attr - config which perf event to use

pid - config which thread to monitor, all threads if -1

cpu - config which cpu to monitor, all cpu is -1

group_fd, flags - usually not used

returns a file descriptor, which can be used to read counter values and records
Kernel support
● pmu drivers
○ register to perf event driver via perf_pmu_register().
○ cpu pmu driver, which operates ARM PMU, lives in drivers/perf.
○ software pmu driver, events like cpu-clock, page-faults, full list is in perf_sw_ids.
○ tracepoint pmu driver, events like sched:sched_switch, full list is in /sys/kernel/tracing/events.
○ device specific pmu drivers.
How simpleperf works

Userspace Kernel Hardware

cpu pmu driver ARM PMU


simpleperf stat cmd

software pmu driver


simpleperf record cmd perf
event tracepoint pmu
driver driver
profile data file
device specific pmu
simpleperf report cmd drivers
and other report
methods
simpleperf commands
● simpleperf is an executable running on device, shipped in /system/bin.
● simpleperf divides its functions into subcommands.
○ list command: list available perf events on device
○ stat command: monitor threads, and print perf event counter values
○ record command: monitor threads, and generate profile data with samples
○ report command: report profile data generated by record command
● simpleperf also provides python scripts running on host
○ to help recording
○ to help reporting
list cmd: list available events
$ simpleperf list

List of hardware events:


branch-misses
bus-cycles
cache-misses
cache-references
cpu-cycles
instructions
stalled-cycles-backend
stalled-cycles-frontend

List of software events:


alignment-faults
context-switches
cpu-clock
…
stat cmd: get perf event counter values
Simpleperf Kernel
Call perf_event_open() for selected
threads, cpus, with selected perf events
Start pmu counters to
count values of the
selected perf events
Wait until stat time ends

Use read() to read counter values from


perf event file descriptors
stat cmd: options
$ simpleperf stat -h

Usage: simpleperf stat [options] [command [command-args]]


Gather performance counter information of running [command].

Options:
-p pid1,pid2,... Stat events on existing processes.
-t tid1,tid2,... Stat events on existing threads.
-a Collect system-wide information.
--cpu cpu_item1,cpu_item2,... Collect information only on the selected cpus.
-e event1[:modifier1],event2[:modifier2],... Select a list of events to count.
--duration time_in_sec Monitor for time_in_sec seconds.
stat cmd: example
$ simpleperf stat -e cache-references,cache-misses -a --duration 1

Performance counter statistics:

# count event_name # count / runtime, runtime / enabled_time


774,728,087 cache-references # 96.513 M/sec (100%)
31,985,983 cache-misses # 4.128672% miss rate (100%)

Total test time: 1.001893 seconds.


record cmd: generate profile data with samples
Simpleperf Kernel

Call perf_event_open() for selected


threads, cpus, with selected perf events Start pmu counter to
count values of the
Use mmap() on perf event fds to create selected pmu event
shared circular buffer between kernel
and simpleperf

Each time the counter


In the main In record read overflows (depending on
thread, process thread, receive the frequency we set by
records, collect records from -f), create a sample
auxiliary circular buffer, record, and put it to the
information and and pass circular buffer
store them in records to the
recording file main thread.
record cmd: options
$ simpleperf record -h

Usage: simpleperf record [options] [--] [command [command-args]]


Gather sampling information of running [command].

Options:
-p pid1,pid2,... Record events on existing processes.
-t tid1,tid2,... Record events on existing threads.
-a System-wide collection.
--cpu cpu_item1,cpu_item2,... Collect information only on the selected cpus.
-e event1[:modifier1],event2[:modifier2],... Select a list of events to count.
-f freq Set event sample frequency. It means recording at
most [freq] samples every second.
--duration time_in_sec Monitor for time_in_sec seconds
-o record_file_name Set record file name, default is perf.data.
--call-graph fp | dwarf[,<dump_stack_size>] Enable call graph recording.
-g Same as '--call-graph dwarf'.
record cmd: example
$ simpleperf record -g sleep 1

simpleperf I cmd_record.cpp:696] Recorded for 1.01908 seconds. Start post processing.


simpleperf I cmd_record.cpp:771] Samples recorded: 56. Samples lost: 0.
record cmd: sample format
The profile data contains a list of samples.

Each sample can contain below information (full list is here):

time - timestamp in CLOCK_MONOTONIC

pid, tid - process id, thread id

cpu - cpu

period - how many events have happened since last sample

ips[] - callstack (frame-pointer based call stack)

regs[] - userspace register values dwarf based call stack generated


stack[] - user stack data up to 64k by stack unwinding
report cmd: report profile data
$ simpleperf report

Cmdline: /system/bin/simpleperf record -g sleep 1


Arch: arm64
Event: cpu-cycles (type 0, config 0)
Samples: 56
Event count: 13885436

Overhead Command Pid Tid Shared Object Symbol


9.61% sleep 14852 14852 [kernel.kallsyms] vma_link
8.97% sleep 14852 14852 linker64 soinfo_do_lookup_impl
6.42% sleep 14852 14852 linker64 BionicAllocator::alloc_impl
6.11% sleep 14852 14852 [kernel.kallsyms] __follow_mount_rcu
5.83% sleep 14852 14852 [kernel.kallsyms] clear_page
…
report profile data on host
Pull record file on host and use multiple report methods (scripts are listed here).
Q&A

You might also like