ThreadMXBean::getCurrentThreadAllocatedBytes

[Hohensee, Paul]

I believe you could move the code ahead of the call to validate_thread_id_array() because that method just checks for thread ids <= 0.


diff -r 3ddf41505d54 src/hotspot/share/services/management.cpp

--- a/src/hotspot/share/services/management.cpp Sun Jun 03 23:33:00 2018 -0700

+++ b/src/hotspot/share/services/management.cpp Mon Jul 16 10:41:28 2018 -0700

@@ -2084,11 +2083,19 @@

   typeArrayOop sa = typeArrayOop(JNIHandles::resolve_non_null(sizeArray));

   typeArrayHandle sizeArray_h(THREAD, sa);



+  // Special-case current thread

+  int num_threads = ids_ah->length();

+  JavaThread* java_thread = JavaThread::current();

+  if (num_threads == 1 && sizeArray_h->length() == 1 &&

+      ids_ah->long_at(0) == java_lang_Thread::thread_id(java_thread->threadObj())) {

+    sizeArray_h->long_at_put(0, java_thread->cooked_allocated_bytes());

+    return;

+  }

+

   // validate the thread id array

   validate_thread_id_array(ids_ah, CHECK);



   // sizeArray must be of the same length as the given array of thread IDs

-  int num_threads = ids_ah->length();

   if (num_threads != sizeArray_h->length()) {

     THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(),

               "The length of the given long array does not match the length of "

If performance is good enough, and if you still want to add getCurrentThreadAllocatedBytes() (imo a good idea, since getCurrentThreadCpuTime() and getCurrentThreadUserTime() already exist), you could implement it by “getThreadAllocatedBytes(Thread::currentThread().getId())”. You might want also want to add getCurrentThread* methods to com.sun.management where they don’t currently exist: then we’d have a complete parallel method set.



Another approach to improving things is to fix the underlying problem with find_JavaThread_from_java_tid(). https://bugs.openjdk.java.net/browse/JDK-8185005 proposes doing that in a different context. We came up with a patch for JDK8 that uses an open addressed hashtable (one where the “bucket chain” is in the index array, see https://en.wikipedia.org/wiki/Hash_table#Open_addressing) to map Java tids to JavaThread*s. I’ve forward ported it to JDK12, see http://cr.openjdk.java.net/~phh/8185005/webrev.00/. The main disadvantage, of course, is that it’s yet another data structure that takes up memory. It’s really fast though and speeds up our profilers quite a bit. Perhaps we could replace the existing thread list with a variation on this map, since it’s quick to just run through the underlying array when you want to run through the threads.



Thanks,



Paul

From: serviceability-dev <serviceability-dev-bounces at openjdk.java.net> on behalf of "Daniel D. Daugherty" <daniel.daugherty at oracle.com>
Reply-To: "daniel.daugherty at oracle.com" <daniel.daugherty at oracle.com>
Date: Friday, July 13, 2018 at 1:53 PM
To: Markus Gaisbauer <markus.gaisbauer at gmail.com>, "serviceability-dev at openjdk.java.net" <serviceability-dev at openjdk.java.net>, Erik Österlund <erik.osterlund at oracle.com>, Robbin Ehn <robbin.ehn at oracle.com>
Subject: Re: ThreadMXBean::getCurrentThreadAllocatedBytes

Markus,

I filed the following bug for you:

    JDK-8207266 ThreadMXBean::getThreadAllocatedBytes() can be quicker for self thread
    https://bugs.openjdk.java.net/browse/JDK-8207266

Dan

On 7/13/18 4:46 PM, Daniel D. Daugherty wrote:
On 7/13/18 2:44 PM, Daniel D. Daugherty wrote:

On 7/13/18 12:35 PM, Markus Gaisbauer wrote:

Hello,

I am trying to use ThreadMXBean::getThreadAllocatedBytes (com.sun.management) to get the amount of allocated memory of the current thread in some performance critical code.

Unfortunately, the current implementation can be rather slow and the duration of each call unpredictable. I ran a test in a JVM with 500 threads. Depending on which thread was queried, getThreadAllocatedBytes took between 100 ns and 2500 ns.

The root cause of the problem is ThreadsList::find_JavaThread_from_java_tid which performs a linear scan through all Java threads in the current process. The more threads a JVM has, the slower it gets. In the worst case, the thread with the given TID is found as the last entry in the list.

Before Java 10, the oldest thread is the slowest one to query.
Since Java 10, the youngest thread is the slowest one to query. I think this was a side effect of introducing "Thread Safe Memory Reclamation (Thread-SMR) support".

             Oldest Thread   Youngest Thread
Java 8             8740 ns             76 ns
Java 10             109 ns           2485 ns

It is good to see that longest search is much faster. Erik and Robbin
will be pleased since speeding up traversal of the ThreadsList was one
of the things that we tried to do during the Thread-SMR project.

A first step is get a new bug filed that documents the issue with
ThreadMXBean::getThreadAllocatedBytes(). Perhaps Gary or Serguei
will take care of that.

Dan



A common use case is to query the metric for the current thread (e.g. before and after performing some operation). This case can be optimized by introducing a new method: getCurrentThreadAllocatedBytes.

I created a patch for http://hg.openjdk.java.net/jdk/jdk/ and by using the new method I saw the following improvements in my test:

             Oldest Thread   Youngest Thread
Proposal             37 ns             37 ns

This is a 60x improvement over the worst case of the current API. In the best case of the current API, the new method is still 3 times faster.

// based on JVM_SetNativeThreadName in jvm.cpp.
JVM_ENTRY(jlong, jmm_GetCurrentThreadAllocatedMemory(JNIEnv *env, jobject currentThread))
  // We don't use a ThreadsListHandle here because the current thread
  // must be alive.
  oop java_thread = JNIHandles::resolve_non_null(currentThread);
  JavaThread* thr = java_lang_Thread::thread(java_thread);
  if (thread == thr) {
    // only supported for the current thread
    return thr->cooked_allocated_bytes();
  }
  return -1;
JVM_END

The proposed method also fixes the problem, that getThreadAllocatedBytes itself allocates some memory on the current thread (two long arrays, 24 bytes) and therefore can slightly skew measurements. The new method, getCurrentThreadAllocatedBytes, returns exactly the same value if it is called twice without allocating any memory between those calls.

I also built a variation of this method that could be used to query allocated memory more efficiently for anyone who already has a java.lang.Thread object:

JVM_ENTRY(jlong, jmm_GetThreadAllocatedMemory(JNIEnv *env, jobject threadObj))
  // based on code proposed in threadSMR.hpp
  ThreadsListHandle tlh;
  JavaThread* thr = NULL;
  bool is_alive = tlh.cv_internal_thread_to_JavaThread(threadObj, &thr, NULL);
  if (is_alive) {
    return thr->cooked_allocated_bytes();
  }
  return -1;
JVM_END

This method took 70 ns in my test, which is 85% slower than GetCurrentThreadAllocatedMemory but still 30% faster than the best case of the current API. I currently have no immediate need for this second method, but I think it would also be a valueable addition to the API.

I attached a patch for getCurrentThreadAllocatedBytes. I can create a second patch for also adding getThreadAllocatedMemory(java.lang.Thread) to the API.

I am a first time contributor and I am not 100% sure what process I must follow to get a change like this into OpenJDK. Can someone have a look at my proposal and help me through the process?

Best regards,
Markus



I believe this is the code that's causing you grief:

open/src/hotspot/share/services/management.cpp:

// Gets an array containing the amount of memory allocated on the Java
// heap for a set of threads (in bytes).  Each element of the array is
// the amount of memory allocated for the thread ID specified in the
// corresponding entry in the given array of thread IDs; or -1 if the
// thread does not exist or has terminated.
JVM_ENTRY(void, jmm_GetThreadAllocatedMemory(JNIEnv *env, jlongArray ids,
                                             jlongArray sizeArray))
  // Check if threads is null
  if (ids == NULL || sizeArray == NULL) {
    THROW(vmSymbols::java_lang_NullPointerException());
  }

  ResourceMark rm(THREAD);
  typeArrayOop ta = typeArrayOop(JNIHandles::resolve_non_null(ids));
  typeArrayHandle ids_ah(THREAD, ta);

  typeArrayOop sa = typeArrayOop(JNIHandles::resolve_non_null(sizeArray));
  typeArrayHandle sizeArray_h(THREAD, sa);

  // validate the thread id array
  validate_thread_id_array(ids_ah, CHECK);

  // sizeArray must be of the same length as the given array of thread IDs
  int num_threads = ids_ah->length();
  if (num_threads != sizeArray_h->length()) {
    THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(),
              "The length of the given long array does not match the length of "
              "the given array of thread IDs");
  }

  ThreadsListHandle tlh;
  for (int i = 0; i < num_threads; i++) {
    JavaThread* java_thread = tlh.list()->find_JavaThread_from_java_tid(ids_ah->long_at(i));
    if (java_thread != NULL) {
      sizeArray_h->long_at_put(i, java_thread->cooked_allocated_bytes());
    }
  }
JVM_END


Perhaps something like this above the "ThreadsListHandle tlh;" line:

  if (num_threads == 1 && THREAD->is_Java_thread()) {
    // Only asking for 1 thread so if we're a JavaThread, then
    // see if this request is for ourself.
    JavaThread* jt = THREAD;
    oop tobj = jt->threadObj();

    if (ids_ah->long_at(0) == java_lang_Thread::thread_id(tobj)) {
      // Return the info for ourself.
      sizeArray_h->long_at_put(0, jt->cooked_allocated_bytes());
      return;
    }
  }

I haven't checked to see if this will even compile, but I
think you'll get the idea.

Dan


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openjdk.java.net/pipermail/serviceability-dev/attachments/20180716/b5d666a9/attachment-0001.html>