Xuhpclab / jxperf

JXPerf

Java inefficiency detection tool based on CPU performance monitoring counters and hardware debug registers. The tool detects dead writes, silent stores, redundant loads, and memory bloat.

Contents

• Installation
• Usage
• Support Platforms
• Support JDK Versions
• License

Installation

Linux

1. Installation Prerequisites

• Install Oracle/OpenJDK and Apache Maven.
• Install libnuma library
• Install python modules: bintrees and google-api-python-client
• Turn on PMU sampling in your environment:sysctl -w kernel.perf_event_paranoid=1
• cp set_env.template set_env
• Modify set_env to make JXPerf_HOME, JAVA_HOME and MAVEN_HOME point to your JXPerf, Java and Maven home.
• source set_env

2. Installation

$ make

3. Uninstallation

$ make clean

Usage

Linux

1. To run dead store detection

• Start Profiler

$ LD_PRELOAD=$JXPerf_HOME/build/preload/libpreload.so java -agentpath:$JXPerf_HOME/build/libagent.so=DeadStore::MEM_UOPS_RETIRED:ALL_STORES:precise=2@<sampling rate> -cp <classpath> <java program>

• Generate profiling results "agent-data"

$ python $JXPerf_HOME/script/process_raw_data.py

2. To run silent store detection

• Start Profiler

$ LD_PRELOAD=$JXPerf_HOME/build/preload/libpreload.so java -agentpath:$JXPerf_HOME/build/libagent.so=SilentStore::MEM_UOPS_RETIRED:ALL_STORES:precise=2@<sampling rate> -cp <classpath> <java program>

• Generate profiling results "agent-data"

$ python $JXPerf_HOME/script/process_raw_data.py

3. To run silent load detection

• Start Profiler

$ LD_PRELOAD=$JXPerf_HOME/build/preload/libpreload.so java -agentpath:$JXPerf_HOME/build/libagent.so=SilentLoad::MEM_UOPS_RETIRED:ALL_LOADS:precise=2@<sampling rate> -cp <classpath> <java program>

• Generate profiling results "agent-data"

$ python $JXPerf_HOME/script/process_raw_data.py

4. To run data centric analysis

• Start Profiler

$ LD_PRELOAD=$JXPerf_HOME/build/preload/libpreload.so java -javaagent:$JAVA_AGENT -agentpath:$JXPerf_HOME/build/libagent.so=DataCentric::MEM_LOAD_UOPS_RETIRED:L1_MISS:precise=2@<sampling rate> -cp <classpath> <java program>

• Generate profiling results "agent-data"

$ python $JXPerf_HOME/script/process_raw_data.py

• The "agent_data" includes two metrics: "Allocation Times" and "L1 Cache Misses"

  • The metric "Allocation Times" reports allocation times for every object, which is represented with the object allocation site

  • The metric "L1 Cache Misses" reports a pair of calling context (i.e., <allocation site, access site>) for every object incurring L1 cache misses

  • To analyze memory bloat

    • Identify the objects suffering from high L1 cache misses by looking into the metric "L1 Cache Misses"
    • Check whether these objects have high allocation times by looking into the metric "Allocation Times"
    • The objects having both high L1 cache misses and allocation times are primary optimization candidates

5. To run NUMA locality analysis

• Start Profiler

$ LD_PRELOAD=$JXPerf_HOME/build/preload/libpreload.so java -javaagent:$JAVA_AGENT -agentpath:$JXPerf_HOME/build/libagent.so=Numa::MEM_LOAD_UOPS_RETIRED:L1_MISS:precise=2@<sampling rate> -cp <classpath> <java program>

• Generate profiling results "agent-data"

$ python $JXPerf_HOME/script/process_raw_data.py

• The "agent_data" includes:
  • "Fraction of Mismatch": mismatch times of one object over the total mismatch times of whole program
  • "Match Times" and "Mismatch Times": the match and mismatch times of one object
  • "Match Percentage": Match Times / (Match Times + Mismatch Times)
  • "Mismatch Percentage": Mismatch Times / (Match Times + Mismatch Times)

6. To run with generic PMU events

• Start Profiler

$ LD_PRELOAD=$JXPerf_HOME/build/preload/libpreload.so java -agentpath:$JXPerf_HOME/build/libagent.so=Generic::PMU_Events:precise=2@<sampling rate> -cp <classpath> <java program>

• Generate profiling results "agent-data"

$ python $JXPerf_HOME/script/process_raw_data.py

7. To run heap profiling

• Start Profiler

$ LD_PRELOAD=$JXPerf_HOME/build/preload/libpreload.so java -agentpath:$JXPerf_HOME/build/libagent.so=Heap -cp <classpath> <java program>

• Generate profiling results "agent-data"

$ python $JXPerf_HOME/script/process_raw_data.py

8. Attach to a running JVM

• Open run_attach.sh and change MODE to one of below modes:
  • DataCentric::MEM_LOAD_UOPS_RETIRED:L1_MISS:precise=2@sampling_rate
  • DeadStore::MEM_UOPS_RETIRED:ALL_STORES:precise=2@sampling_rate
  • SilentStore::MEM_UOPS_RETIRED:ALL_STORES:precise=2@sampling_rate
  • SilentLoad::MEM_UOPS_RETIRED:ALL_LOADS:precise=2@sampling_rate
• Start Profiler

$ ./run_attach.sh <running time in seconds> <pid>

VS Code GUI

• In VS Code, search for DrCCTProf Viewer extension and install it
• Generate the drcctprof format profile (test.drcctprof):$JXPerf_HOME/script/process_raw_data_to_vscode.py <source code foler>
• View the test.drcctprof in VS Code:code test.drcctprof

Support Platforms

We tested our tool on following platforms.

Linux

CPU	Systems	Kernel	Architecture
Intel(R) Xeon(R) CPU E5-2650 v4	Ubuntu 14.04.6	Linux 5.1.0	x86_64
Intel(R) Xeon(R) CPU E5-2699 v3	Ubuntu 18.04.3	Linux 5.4.6	x86_64
Intel(R) Xeon(R) CPU E7-4830 v4	CentOS Linux 7	Linux 3.10.0	x86_64

Support JDK Versions

JDK

JDK Versions
JDK 11 & later
OpenJDK 11 & later

License

JXPerf is released under the MIT License.