/jvm-profiling-toolkit

Primary LanguageJavaApache License 2.0Apache-2.0

JVM profiling toolkit

Viewer for collapsed stack and JFR outputs of profiles. Dedicated to Async-profiler 2.x, but works also with Async-profiler 1.x and Java Flight Recorder.

ECID functionality

To benefit from filters with ECID you need to use precompiled Async-profiler from PR, and make use of a new API created there.

JAVA_HOME should point to JDK 17.

How to install - from sources

git clone --depth 1 https://github.com/krzysztofslusarski/jvm-profiling-toolkit.git
cd jvm-profiling-toolkit/
mvn clean package

How to run

java -jar viewer-application/target/viewer-application.jar

Java should point to JDK 11. After you run it, viewer is available on http://localhost:8079/.

How to configure

Viewer is Spring Boot application, you can create application.yml with:

server:
  port: 8079

Example usage od Async-profiler for collapsed stack

./profiler -t -d 30 -e cpu -o collapsed -f output.txt <pid>

  • -t - gives you output divided by thread
  • -d 30 - 30s duration
  • -e cpu - profiled event, viewer should work with every event
  • -o collapsed - as name suggest, this is collapsed stack viewer, so this is mandatory output
  • -f output.txt - output file
  • <pid> - pid of your JVM

Features of viewer

Analysis of collapsed stack file

Flame graphs

Flame graphs

Viewer can generate flame graphs:

  • Flame graph - common flame graph from your collapsed stack file
  • Flame graph with no thread division - common flame graph with division by thread removed
  • Hotspot flame graph - flame graph that is inverted and reversed, presenting hotspots from collapsed stack file:
    • depth = 10/20/30 - shortened graph with smaller stacks

Method total time

Total time

Total time is number of stacks, that method was anywhere on the stack. Method name can be filtered.

Method self time

Self time

Self time is number of stacks, that method was at the end of the stack. Method name can be filtered.

Callee and callers flame graphs for methods

Callee

Callee graph shows what method is actually doing. This graph is aggregated, so it shows every usage of method.

Callee

Callers graph shows what which method used profiled method. This graph is aggregated, so it shows every usage of method.

Analysis of JFR files

You can add filters to your parser:

  • Thread filter
  • Access log style filter - end date and duration (in milliseconds)
  • Warmup / cooldown filter - this one skips proper number of seconds from the beginning and the end

Collapsed stack files

You can upload multiple JFR files with a single HTTP POST to the analyzer. Analyzer creates following collapsed stack from your JFR file:

  • Wall-clock - if you used Async-profiler in wall mode only
  • CPU - if you used Async-profiler in wall mode the CPU file is made from wall output with stacks that were consuming CPU only
  • Heap allocation (count) - if you used alloc mode - this one presents count of allocations that needed new TLAB or needed allocation outside TLAB
  • Heap allocation (size) - if you used alloc mode - this one presents size of allocations mentioned above
  • Locks - if you used lock mode
  • CPU load
    • JVM system
    • JVM user
    • JVM total
    • Machine total
    • Machine total - JVM total
CPU usage flame graphs

Cpu usage

On example above you can see that the CPU (this one is JVM total) consumed at least 30% of the CPU for 71,21% of time.

Other JFR information

JFR viewer will also show you:

  • OS Info
  • CPU Info
  • Initial system properties
  • JVM Info

If you upload multiple files then last information parsed is present in those sections.