Async-profiler: ReentrantLock contention

[franz1981]

Hi guys,

First on above: thanks for this awesome project and for the hard work on it!!!

Now my question: would be possible to have a contention profiling for Java's ReentrantLocks similar to the -e lock option?

I was looking http://www.brendangregg.com/offcpuanalysis.html to find something similar, given that under contention ReentrantLock uses Unsafe::park/unpark, but having it on the async-profiler toolkit would be awesome.
If there is something I could help to make it possible I would be very happy to contribute :+1:

Thanks,
Franz

[apangin]

Yes, this is absolutely possible. I think I can implement this if not in the nearest, then in the next release.

[franz1981]

That's would be awesome, thanks!

[apangin]

-e lock now handles both intrinsic locks and ReentrantLocks.

[franz1981]

Thanks for this change!!!
I've just refreshed by local repo with the last changes and compiled the whole with make.
I'm not a C guy and I'm hoping to do not have mess with anything...

I've built this test program:

   public static void main(String[] args) {
      final int tokens = 1;
      final boolean jucLock = true;
      final ReentrantLock lock = new ReentrantLock();
      final Object obj = new Object();
      final int writers = 4;
      final AtomicBoolean stop = new AtomicBoolean();
      ExecutorService executorService = Executors.newCachedThreadPool();
      for (int i = 0; i < writers; i++) {
         if (jucLock) {
            executorService.execute(() -> {
               while (!stop.get()) {
                  lock.lock();
                  Blackhole.consumeCPU(tokens);
                  lock.unlock();
               }
            });
         } else {
            executorService.execute(() -> {
               while (!stop.get()) {
                  synchronized (obj) {
                     Blackhole.consumeCPU(tokens);
                  }
               }
            });
         }
      }
      executorService.shutdown();
      Runtime.getRuntime().addShutdownHook(new Thread(() -> {
         stop.lazySet(true);
      }));
   }

Using synchronized I'm getting this:

While for the ReentrantLock I'm receiving from flamegraph:

ERROR: No stack counts found

Probably I've done something wrong on compilation or the test I've written has something wrong.
There is anything I could do to help beside looking at make output (that doesn't seems to have emitted any warning/error)?

[apangin]

Works for me. How do you run the target program and the profiler? What JDK version do you use?

./profiler.sh -d 10 -e lock -f locks.svg PID generates the following flame graph:

[franz1981]

I have finally managed to understand the error, sorry :(
I wasn't running correctly the program :O
Now I finally see them:

What is strange is that now I can see only ReentrantLock(s) but no intrinsic locks, maybe is just a case.
With the previous version of the profiler I remember that there were intrinsic locks too: I will run a program that will use both to see if is true :+1:

[apangin]

I guess in your case the number of contended intrinsic locks is so small comparing to ReentrantLocks, that intrinsic locks were skipped as negligible from the result flamegraph.

You may modify the threshold with --minwidth argument or try different output format (e.g. -o traces) to make sure intrinsic locks are still there.

[franz1981]

I guess in your case the number of contended intrinsic locks is so small comparing to ReentrantLocks, that intrinsic locks were skipped as negligible from the result flamegraph.

Guess correct!
Thanks to have added this feature in such a small time! :+1: