Table of Contents


profiler online tutorialopen in new window


Generate a flame graph using async-profileropen in new window

The profiler command supports generating flame graph for application hotspots.

The basic usage of the profiler command is profiler action [actionArg]

The arguments of profiler command basically keeps consistent with upstream project async-profileropen in new window, you can refer to its README, Github Discussions and other documentations for further information of usage.

Supported Options

actionAction to execute
actionArgAttribute name pattern
[i:]sampling interval in ns (default: 10'000'000, i.e. 10 ms)
[f:]dump output to specified directory
[d:]run profiling for specified seconds
[e:]which event to trace (cpu, alloc, lock, cache-misses etc.), default value is cpu

Start profiler

$ profiler start
Started [cpu] profiling


By default, the sample event is cpu. Other valid profiling modes can be specified with the --event parameter, see relevant contents below.

Get the number of samples collected

$ profiler getSamples

View profiling status

$ profiler status
[cpu] profiling is running for 4 seconds

Can view which event and sampling time.

View profiler memory usage

$ profiler meminfo
Call trace storage:   10244 KB
      Dictionaries:      72 KB
        Code cache:   12890 KB
             Total:   23206 KB

Stop profiler

Generating flame graph results

By default, the result file is html file in Flame Graphopen in new window format. You can also specify other format with the -o or --format parameter, including flat, traces, collapsed, flamegraph, tree, jfr:

$ profiler stop --format flamegraph
profiler output file: /tmp/test/arthas-output/20211207-111550.html

When extension of filename in --file parameter is html or jfr, the output format can be infered. For example, --file /tmp/result.html will generate flamegraph automatically.

View profiler results under arthas-output via browser

By default, arthas uses port 3658, which can be opened: http://localhost:3658/arthas-output/open in new window View the arthas-output directory below Profiler results:

Click to view specific results:


If using the chrome browser, may need to be refreshed multiple times.

Profiler supported events

Under different platforms and different OSs, the supported events are different. For example, under macos:

$ profiler list
Basic events:

Under linux

$ profiler list
Basic events:
Java method calls:
Perf events:

If you encounter the permissions/configuration issues of the OS itself and then missing some events, you can refer to the async-profileropen in new window documentation.

You can use check action to check if a profiling event is available, this action receives the same format options with start.

You can use the --event parameter to specify the event to sample, for example, alloc event means heap memory allocation profiling:

$ profiler start --event alloc

Resume sampling

$ profiler resume
Started [cpu] profiling

The difference between start and resume is: start will clean existing result of last profiling before starting, resume will retain the existing result and add result of this time to it.

You can verify the number of samples by executing profiler getSamples.

Dump action

$ profiler dump

The dump action saves profiling result to default file or specified file, but profiling will continue. That means if you start profiling and dump after 5 seconds, then dump after 2 seconds again, you will get 2 result files, the first one contains profiling result of 0~5 seconds and the second one contains that of 0~7 seconds.

Use execute action to execute complex commands

For example, start sampling:

profiler execute 'start,framebuf=5000000'

Stop sampling and save to the specified file:

profiler execute 'stop,file=/tmp/result.html'

Specific format reference: arguments.cppopen in new window

View all supported actions

$ profiler actions
Supported Actions: [resume, dumpCollapsed, getSamples, start, list, version, execute, meminfo, stop, load, dumpFlat, dump, actions, dumpTraces, status, check]

View version

$ profiler version
Async-profiler 2.9 built on May  8 2023
Copyright 2016-2021 Andrei Pangin

Configure Java stack depth

You can use -j or --jstackdepth option to configure maximum Java stack depth. This option will be ignored if value is greater than default 2048. This option is useful when you don't want to see stacks that are too deep. Below is usage example:

profiler start -j 256

Profiling different threads separately

You can use -t or --threads flag option to profile different threads separately, each stack trace will end with a frame that denotes a single thread.

profiler start -t

Configure include/exclude to filter data

If the application is complex and generates a lot of content, and you want to focus on only part of stack traces, you can filter stack traces by --include/--exclude. --include defines the name pattern that must be present in the stack traces, while --exclude is the pattern that must not occur in any of stack traces in the output.A pattern may begin or end with a star * that denotes any (possibly empty) sequence of characters. such as

profiler stop --include'java/*' --include 'com/demo/*' --exclude'*Unsafe.park*'

Both --include/--exclude support being set multiple times, but need to be configured at the end of the command line. You can also use short parameter format -I/-X. Note that --include/--exclude only supports configuration at stop action or start action with -d/--duration parameter, otherwise it will not take effect.

Specify execution time

For example, if you want the profiler to automatically end after 300 seconds, you can specify it with the -d/--duration parameter in collect action:

profiler collect --duration 300

Generate jfr format result

Note that jfr only supports configuration at start. If it is specified at stop, it will not take effect.

profiler start --file /tmp/test.jfr
profiler start -o jfr

The file parameter supports some variables:

  • Timestamp: --file /tmp/test-%t.jfr
  • Process ID: --file /tmp/test-%p.jfr

The generated results can be viewed with tools that support the jfr format. such as:

  • JDK Mission Control:
  • JProfiler:

Control details in result

The -s parameter will use simple name instead of Fully qualified name, e.g. MathGame.main instead of demo.MathGame.main. The -g parameter will use method signatures instead of method names, e.g. demo.MathGame.main([Ljava/lang/String;)V instead of demo.MathGame.main. There are many parameters related to result format details, you can refer to async-profiler READMEopen in new window and async-profiler Github Discussionsopen in new window and other information.

For example, in command below, -s use simple name for Java class, -g show method signatures, -a will annotate Java methods, -l will prepend library names for native method, --title specify a title for flame graph page, --minwidth will skip frames smaller than 15% in flame graph, --reverse will generate stack-reversed FlameGraph / Call tree.

profiler stop -s -g -a -l --title <flametitle> --minwidth 15 --reverse

The 'unknown' in profiler result


Config locks/allocations profiling threshold

When profiling in locks or allocations event, you can use --lock or --alloc to config thresholds, for example:

profiler start -e lock --lock 10ms
profiler start -e alloc --alloc 2m

will profile contended locks longer than 10ms (default unit is ns if no unit is specified), or profile allocations with 2m BYTES interval.

Config JFR chunks

When using JFR as output format, you can use --chunksize or --chunktime to config approximate size (in bytes, default value is 100MB) and time limits (default value is 1 hour) for a single JFR chunk. For example:

profiler start -f profile.jfr --chunksize 100m --chunktime 1h

Group threads by scheduling policy

You can use --sched flag option to group threads in output by Linux-specific scheduling policy: BATCH/IDLE/OTHER, for example:

profiler start --sched

The second line from bottom in flamegraph represent the scheduling policy.

Build allocation profile from live objects only

Use --live flag option to retain allocation samples with live objects only (object that have not been collected by the end of profiling session). Useful for finding Java heap memory leaks.

profiler start --live

Config method of collecting C stack frames

Use --cstack MODE to config how to walk native frames (C stack). Possible modes are fp (Frame Pointer), dwarf (DWARF unwind info), lbr (Last Branch Record, available on Haswell since Linux 4.1), and no (do not collect C stack).

By default, C stack is shown in cpu, itimer, wall-clock and perf-events profiles. Java-level events like alloc and lock collect only Java stack.

profiler --cstack fp

The command above will collection Frame Pointer of C stacks.

Begin or end profiling when FUNCTION is executed

Use --begin function and --end function to automatically start/stop profiling when the specified native function is executed. Its main purpose is to profile certain JVM phases like GC and Safepoint pauses. You should use native function name defined in a JVM implement, for example SafepointSynchronize::begin and SafepointSynchronize::end in HotSpot JVM.

Time-to-safepoint profiling

The --ttsp option is an alias for --begin SafepointSynchronize::begin --end RuntimeService::record_safepoint_synchronized. It is not a separate event type, but rather a constraint. Whatever event type you choose (e.g. cpu or wall), the profiler will work as usual, except that only events between the safepoint request and the start of the VM operation will be recorded.

profiler start --begin SafepointSynchronize::begin --end RuntimeService::record_safepoint_synchronized
profiler --ttsp

Use events from profiler for Java Flight Recording

Use --jfrsync CONFIG to start Java Flight Recording with the given configuration synchronously with the profiler. The output .jfr file will include all regular JFR events, except that execution samples will be obtained from async-profiler. This option implies -o jfr.

CONFIG can be profile, means using the predefined JFR config "profile" in $JAVA_HOME/lib/jfr/, or full path of a JFR configuration file (.jfc), this value has the same format with settings option of JFR.startopen in new window.

For example, command below use "profile" config of JFR:

profiler start -e cpu --jfrsync profile -f combined.jfr

Run profiler in a loop

Use --loop TIME to run profiler in a loop (continuous profiling). The argument is either a clock time (hh:mm:ss) or a loop duration in seconds, minutes, hours, or days. Make sure the filename includes a timestamp pattern, or the output will be overwritten on each iteration. The command below will run profiling endlessly and save records of each hour to a jfr file.

profiler start --loop 1h -f /var/log/profile-%t.jfr

--timeout option

This option specifies the time when profiling will automatically stop. The format is the same as in loop: it is either a wall clock time (12:34:56) or a relative time interval (2h).

Both --loop and --timeout are used for start action but not for collect action, for further information refer to async-profiler Github Discussionsopen in new window.