async-profiler

This project is a low overhead sampling profiler for Java that does not suffer from Safepoint bias problem. It features HotSpot-specific APIs to collect stack traces and to track memory allocations. The profiler works with OpenJDK, Oracle JDK and other Java runtimes based on HotSpot JVM.

async-profiler can trace the following kinds of events:

CPU cycles
Hardware and Software performance counters like cache misses, branch misses, page faults, context switches etc.
Allocations in Java Heap
Contented lock attempts of Java monitors

CPU profiling

In this mode profiler collects stack trace samples that include Java methods, native calls, JVM code and kernel functions.

The general approach is receiving call stacks generated by perf_events and matching them up with call stacks generated by AsyncGetCallTrace, in order to produce an accurate profile of both Java and native code. Additionally, async-profiler provides a workaround to recover stack traces in some corner cases where AsyncGetCallTrace fails.

This approach has the following advantages compared to using perf_events directly with a Java agent that translates addresses to Java method names:

Works on older Java versions because it doesn't require -XX:+PreserveFramePointer, which is only available in JDK 8u60 and later.
Does not introduce the performance overhead from -XX:+PreserveFramePointer, which can in rare cases be as high as 10%.
Does not require generating a map file to map Java code addresses to method names.
Works with interpreter frames.
Does not require writing out a perf.data file for further processing in user space scripts.

ALLOCATION profiling

Instead of detecting CPU-consuming code, the profiler can be configured to collect call sites where the largest amount of heap memory is allocated.

async-profiler does not use intrusive techniques like bytecode instrumentation or expensive DTrace probes which have significant performance impact. It also does not affect Escape Analysis or prevent from JIT optimizations like allocation elimination. Only actual heap allocations are measured.

The profiler features TLAB-driven sampling. It relies on HotSpot-specific callbacks to receive two kinds of notifications:

when an object is allocated in a newly created TLAB;
when an object is allocated on a slow path outside TLAB.

This means not each allocation is counted, but only allocations every N kB, where N is the average size of TLAB. This makes heap sampling very cheap and suitable for production. On the other hand, the collected data may be incomplete, though in practice it will often reflect the top allocation sources.

Unlike Java Mission Control which uses similar approach, async-profiler does not require Java Flight Recorder or any other JDK commercial feature. It is completely based on open source technologies and it works with OpenJDK.

The minimum supported JDK version is 7u40 where the TLAB callbacks appeared.

Heap profiler requires HotSpot debug symbols. Oracle JDK already has them embedded in libjvm.so, but in OpenJDK builds they are typically shipped in a separate package. For example, to install OpenJDK debug symbols on Debian / Ubuntu, run

# apt-get install openjdk-8-dbg

Supported platforms

Linux / x64 / x86 / ARM / AArch64
macOS / x64

Note: macOS profiling is limited only to Java code, since native stack walking relies on perf_events API which is available only on Linux platforms.

Building

Build status:

Make sure the JAVA_HOME environment variable points to your JDK installation, and then run make. GCC is required. After building, the profiler agent binary will be in the build subdirectory. Additionally, a small application jattach that can load the agent into the target process will also be compiled to the build subdirectory.

Basic Usage

As of Linux 4.6, capturing kernel call stacks using perf_events from a non- root process requires setting two runtime variables. You can set them using sysctl or as follows:

# echo 1 > /proc/sys/kernel/perf_event_paranoid
# echo 0 > /proc/sys/kernel/kptr_restrict

To run the agent and pass commands to it, the helper script profiler.sh is provided. A typical workflow would be to launch your Java application, attach the agent and start profiling, exercise your performance scenario, and then stop profiling. The agent's output, including the profiling results, will be displayed in the Java application's standard output.

Example:

$ jps
9234 Jps
8983 Computey
$ ./profiler.sh start 8983
$ ./profiler.sh stop 8983

Alternatively, you may specify -d (duration) argument to profile the application for a fixed period of time with a single command.

$ ./profiler.sh -d 30 8983

By default, the profiling frequency is 1000Hz (every 1ms of CPU time). Here is a sample of the output printed to the Java application's terminal:

--- Execution profile ---
Total:                   687
Unknown (native):        1 (0.15%)

Samples: 679 (98.84%)
    [ 0] Primes.isPrime
    [ 1] Primes.primesThread
    [ 2] Primes.access$000
    [ 3] Primes$1.run
    [ 4] java.lang.Thread.run

... a lot of output omitted for brevity ...

         679 (98.84%) Primes.isPrime
           4 (0.58%)  __do_softirq

... more output omitted ...

This indicates that the hottest method was Primes.isPrime, and the hottest call stack leading to it comes from Primes.primesThread.

Flame Graph visualization

async-profiler provides out-of-the-box Flame Graph support. Specify -o svg argument to dump profiling results as an interactive SVG immediately viewable in all mainstream browsers. Also, SVG output format will be chosen automatically if the target filename ends with .svg.

$ jps
9234 Jps
8983 Computey
$ ./profiler.sh -d 30 -f /tmp/flamegraph.svg 8983

Profiler Options

The following is a complete list of the command-line options accepted by profiler.sh script.

start - starts profiling in semi-automatic mode, i.e. profiler will run until stop command is explicitly called.
stop - stops profiling and prints the report.
status - prints profiling status: whether profiler is active and for how long.
list - show the list of available profiling events. This option still requires PID, since supported events may differ depending on JVM version.
-d N - the profiling duration, in seconds. If no start, stop or status option is given, the profiler will run for the specified period of time and then automatically stop.
Example: ./profiler.sh -d 30 8983
-e event - the profiling event: cpu, alloc, lock, cache-misses etc. Use list to see the complete list of available events.

In allocation profiling mode the top frame of every call trace is the class of the allocated object, and the counter is the total allocated bytes in all samples of the given call trace.

In lock profiling mode the top frame is the class of monitor object, and the counter is number of nanoseconds it took to enter the monitor.
-i N - sets the profiling interval, in nanoseconds. Only CPU active time is counted. No samples are collected while CPU is idle. The default is 1000000 (1ms).
Example: ./profiler.sh -i 100000 8983
-b N - sets the frame buffer size, in the number of Java method ids that should fit in the buffer. If you receive messages about an insufficient frame buffer size, increase this value from the default.
Example: ./profiler.sh -b 5000000 8983
-t - profile threads separately. Each stack trace will end with a frame that denotes a single thread. Example: ./profiler.sh -t 8983
-o fmt[,fmt...] - specifies what information to dump when profiling ends. This is a comma-separated list of the following options:
- summary - dump basic profiling statistics;
- traces[=N] - dump call traces (at most N samples);
- flat[=N] - dump flat profile (top N hot methods);
- collapsed[=C] - dump collapsed call traces in the format used by FlameGraph script. This is a collection of call stacks, where each line is a semicolon separated list of frames followed by a counter.
- svg[=C] - produce Flame Graph in SVG format.
C is a counter type:
- samples - the counter is a number of samples for the given trace;
- total - the counter is a total value of collected metric, e.g. total allocation size.
The default format is summary,traces=200,flat=200.
-f FILENAME - the file name to dump the profile information to.
Example: ./profiler.sh -o collapsed -f /tmp/traces.txt 8983

Restrictions/Limitations

On most Linux systems, perf_events captures call stacks with a maximum depth of 127 frames. On recent Linux kernels, this can be configured using sysctl kernel.perf_event_max_stack or by writing to the /proc/sys/kernel/perf_event_max_stack file.
Profiler allocates 8kB perf_event buffer for each thread of the target process. Make sure /proc/sys/kernel/perf_event_mlock_kb value is large enough (more than 8 * threads) when running under unprivileged user. Otherwise the message "perf_event mmap failed: Operation not permitted" will be printed, and no native stack traces will be collected.
There is no bullet-proof guarantee that the perf_events overflow signal is delivered to the Java thread in a way that guarantees no other code has run, which means that in some rare cases, the captured Java stack might not match the captured native (user+kernel) stack.
You will not see the non-Java frames preceding the Java frames on the stack. For example, if start_thread called JavaMain and then your Java code started running, you will not see the first two frames in the resulting stack. On the other hand, you will see non-Java frames (user and kernel) invoked by your Java code.

Troubleshooting

Could not start attach mechanism: No such file or directory means that the profiler cannot establish communication with the target JVM through UNIX domain socket.

For the profiler to be able to access JVM, make sure

You run profiler under exactly the same user as the owner of target JVM process.
/tmp directory of Java process is physically the same directory as /tmp of your shell.
JVM is not run with -XX:+DisableAttachMechanism option.

[frame_buffer_overflow] in the output means there was not enough space to store all call traces. Consider increasing frame buffer size with -b option.

SamSharaf/async-profiler