chappie

chappie is a statistical profiler that estimates component and method energy consumption for Java runtimes on linux. chappie has only been tested on debian 9.

building

chappie requires the following build tools:

• java (works on 8 but is intended for use with 9+)
• ant
• maven
• jni
• make

to build all necessary components, first build the submodules in vendor. Then run:

ant deps
ant jar

ant will take care of the rest.

profiling

chappie is run asynchronously around a block of code:

import chappie.Chaperone;

public class FooWrapper {
  public static void main() {
    Chaperone chappie = new Chaperone();
    chappie.start();

    // program body

    chappie.stop();
  }
}

running

chappie.sh mimics java calls but also sets up dependencies:

./chappie.sh -cp foo-wrapper.jar FooWrapper foo_0

chappie args

chappie accepts the following arguments:

• -d/--work-directory: path where collected data is stored; defaults to ./chappie-logs
• -Dchappie.<arg>: chappie uses system properties to manually control sampling rates
  • rate: base sampling rate in millis; rate of 0 will only collect CPU frequency data at 512ms
  • vm/os/freq/rapl: sampling factor for profilers; the profiler will collect data every -Dchappie.rate * -Dchappie.<profiler_rate> ms
  • trace: sampling rate for the async-profiler in ns; the buffers are flushed every 10 epochs, so make sure the async-profiler rate is not too fast relative to the base rate

example:

# profile FooWrapper at 20ms for all profilers except the os profiler at 80ms
./chappie.sh -d ~/data/foo-experiment -Dchappie.rate=20 -Dchappie.os=4 -cp foo-wrapper.jar FooWrapper foo_0