RFR: 8185005: Improve performance of ThreadMXBean.getThreadInfo(long ids[], int maxDepth)

classic Classic list List threaded Threaded
3 messages Options
Reply | Threaded
Open this post in threaded view
|

RFR: 8185005: Improve performance of ThreadMXBean.getThreadInfo(long ids[], int maxDepth)

Daniil Titov
Please review the change that improves performance of ThreadMXBean MXBean methods returning the
information for specific threads. The change introduces the thread table that uses ConcurrentHashTable
to store one-to-one the mapping between the thread ids and JavaThread objects and replaces the linear
search over the thread list in ThreadsList::find_JavaThread_from_java_tid(jlong tid) method with the lookup
in the thread table.

Testing: Mach5 tier1,tier2 and tier3 tests successfully passed.

Webrev: https://cr.openjdk.java.net/~dtitov/8185005/webrev.01/ 
Bug: https://bugs.openjdk.java.net/browse/JDK-8185005 

Thanks!

Best regards,
Daniil



Reply | Threaded
Open this post in threaded view
|

Re: 8185005: Improve performance of ThreadMXBean.getThreadInfo(long ids[], int maxDepth)

David Holmes
Hi Daniil,

On 13/08/2019 9:24 am, Daniil Titov wrote:
> Hi Robbin,
>
> Thank you very much for reviewing this version of the fix!  Based on your findings
> it seems as it makes sense to make a step back and  continue with the
> approach we took before in the previous version of the webrev (webrev.04),
> and get more information about the impact on the startup time it has. I will
> consult with Claus regarding this and then share the findings.

That seems a good approach to me. It wasn't at all clear to me that the
latest proposed approach would actually solve the original problem in a
satisfactory way - it would depend on how constant the set of threads
being queried was.

There is no perfect solution here as any fix to the reported problem
incurs overhead elsewhere. Even evaluating the merits of the different
trade-offs is hard to do - we could end up with a compromise solution
that fails to satisfy anyone.

David
-----

> Thanks again,
> --Daniil
>
>
>
>
>
>
> On 8/12/19, 5:22 AM, "Robbin Ehn" <[hidden email]> wrote:
>
>      Hi Daniil,
>      
>      I took a new deeper dive into this.
>      
>      This line seems to have some issues:
>      
>      if (ThreadTable::is_initialized() && thread->in_thread_table() &&
>      !thread->is_attaching_via_jni()) {
>      
>      If you create new threads which attaches and then dies, the table will just keep
>      growing. So you must remove them also ?
>      
>      Secondly you should not use volatile semantics for _in_thread_table.
>      The load in the if-statement can be reordered with _is_initialized.
>      Which could lead to a leak, rogue pointer in the table.
>      
>      So both "static volatile bool _is_initialized;" and "volatile bool
>      _in_thread_table; "
>      should be stored with store_release and loaded with load_acquire.
>      
>      Unfortunately it looks like there still would be races if
>      ThreadTable::add_thread e.g. context switch at:
>      
>      if (_local_table->insert(thread, lookup, entry)) {
>      // HERE
>         java_thread->set_in_thread_table(true);
>      
>      *Remove side can pass the if-statement without removing.
>      
>      Since this thread also maybe exiting at any moment, e.g. context switch:
>      
>             if (tobj != NULL && !thread->is_exiting() &&
>                 java_tid == java_lang_Thread::thread_id(tobj)) {
>       // HERE
>               ThreadTable::add_thread(java_tid, thread);
>      
>      *Add side can add a thread that is exiting.
>      
>      Mixing in a third thread looking up a random tid and getting a JavaThread*, it
>      must validate it against it's ThreadsList. Making the hashtable useless.
>      
>      So I think the only one adding and removing should be the thread itself.
>      1:Add to ThreadsList
>      2:Add to ThreadTable
>      3:Remove from ThreadTable
>      4:Remove ThreadsList
>      
>      Between 1-2 and 3-4 the thread would be looked-up via linear scan.
>      I don't see an easy way around the start-up issue with this.
>      
>      Maybe have the cache in Java.
>      Pass in the thread obj into a
>      java_sun_management_ThreadImpl_getThreadTotalCpuTime3 instead,
>      thus skipping any look-ups in native.
>      
>      Thanks, Robbin
>      
>      
>      On 8/12/19 5:49 AM, Daniil Titov wrote:
>      > Hi David, Robbin, Daniel, and Serguei,
>      >
>      > Please review a new version of the fix.
>      >
>      > As David suggested I created a separated Jira issue [1] to cover  additional optimization for
>      > some callers of find_JavaThread_from_java_tid() and this version of the fix no longer includes
>      > changes in management.cpp ( and the test related with these changes).
>      >
>      > Regarding the impact the previous version of the fix had on the thread startup time at heavy load (e.g.
>      > when 5000 threads are created and destroyed every second) I tried a different approach that makes
>      > calls to ThreadTable::add_thread  and ThreadTable::remove_thread  asynchronous and offloads the
>      > work for actual modifications of the thread table to a periodic task that runs every 5 seconds. With the
>      > same  stress test scenario (the  test does some warm-up and then measures the time it takes to create
>      > and start 100,000 threads; every  thread just sleeps  for 100 ms) the impact on the thread startup time
>      > was reduced to 1.2% ( from 2.7%).
>      >
>      > The cause of this impact in this stress test scenario is that as soon as the thread table is initialized,
>      > an additional work to insert  and delete entries in the thread table should be performed, even if
>      > com.sun.management.ThreadMXBean methods are no longer called. For example, In the stress test
>      > mentioned above, every second about 5000 entries had to be inserted in the table and then deleted.
>      >
>      > That doesn't look right and the new version of the fix uses the different approach: the thread is added to
>      > the thread table only when this thread is requested by com.sun.management.ThreadMXBean bean. Every
>      > time when find_JavaThread_from_java_tid() is called for a new tid, the thread  is found by the iterating over
>      > the thread list and added to the thread table. All consequent calls to find_JavaThread_from_java_tid() for
>      > the same tid returns the thread from the thread table.
>      >
>      > Running stress test for the cases when the thread table is enabled and not showed no difference in the
>      > average thread startup times.
>      >
>      > [1] : https://bugs.openjdk.java.net/browse/JDK-8229391
>      >
>      > Bug: https://bugs.openjdk.java.net/browse/JDK-8185005
>      > Webrev: https://cr.openjdk.java.net/~dtitov/8185005/webrev.05/
>      >
>      > Thanks,
>      > Daniil
>      >
>      > On 8/4/19, 7:54 PM, "David Holmes" <[hidden email]> wrote:
>      >
>      >      Hi Daniil,
>      >
>      >      On 3/08/2019 8:16 am, Daniil Titov wrote:
>      >      > Hi David,
>      >      >
>      >      > Thank you for your detailed review. Please review a new version of the fix that includes
>      >      > the changes you suggested:
>      >      > - ThreadTableCreate_lock scope is reduced to cover the creation of the table only;
>      >      > - ThreadTableCreate_lock is made _safepoint_check_always;
>      >
>      >      Okay.
>      >
>      >      > - ServiceThread is no longer responsible for the resizing of the thread table, instead,
>      >      >    the thread table is changed to grow on demand by the thread that is doing the addition;
>      >
>      >      Okay - I'm happy to get the serviceThread out of the picture here.
>      >
>      >      > - fixed nits and formatting issues.
>      >
>      >      Okay.
>      >
>      >      >>> The change also includes additional optimization for some callers of find_JavaThread_from_java_tid()
>      >      >>>   as Daniel suggested.
>      >      >> Not sure it's best to combine these, but if they are limited to the
>      >      >> changes in management.cpp only then that may be okay.
>      >      >
>      >      > The additional optimization for some callers of find_JavaThread_from_java_tid() is
>      >      > limited to management.cpp (plus a new test) so I left them in the webrev  but
>      >      > I also could move it in the separate issue if required.
>      >
>      >      I'd prefer this part of be separated out, but won't insist. Let's see if
>      >      Dan or Serguei have a strong opinion.
>      >
>      >      >    > src/hotspot/share/runtime/threadSMR.cpp
>      >      >    >755     jlong tid = SharedRuntime::get_java_tid(thread);
>      >      >    > 926     jlong tid = SharedRuntime::get_java_tid(thread);
>      >      >   >  I think it cleaner/better to just use
>      >      >   > jlong tid = java_lang_Thread::thread_id(thread->threadObj());
>      >      >   > as we know thread is not NULL, it is a JavaThread and it has to have a
>      >      >   > non-null threadObj.
>      >      >
>      >      > I had to leave this code unchanged since it turned out the threadObj is null
>      >      > when VM is destroyed:
>      >      >
>      >      > V  [libjvm.so+0xe165d7]  oopDesc::long_field(int) const+0x67
>      >      > V  [libjvm.so+0x16e06c6]  ThreadsSMRSupport::add_thread(JavaThread*)+0x116
>      >      > V  [libjvm.so+0x16d1302]  Threads::add(JavaThread*, bool)+0x82
>      >      > V  [libjvm.so+0xef8369]  attach_current_thread.part.197+0xc9
>      >      > V  [libjvm.so+0xec136c]  jni_DestroyJavaVM+0x6c
>      >      > C  [libjli.so+0x4333]  JavaMain+0x2c3
>      >      > C  [libjli.so+0x8159]  ThreadJavaMain+0x9
>      >
>      >      This is actually nothing to do with the VM being destroyed, but is an
>      >      issue with JNI_AttachCurrentThread and its interaction with the
>      >      ThreadSMR iterators. The attach process is:
>      >      - create JavaThread
>      >      - mark as "is attaching via jni"
>      >      - add to ThreadsList
>      >      - create java.lang.Thread object (you can only execute Java code after
>      >      you are attached)
>      >      - mark as "attach completed"
>      >
>      >      So while a thread "is attaching" it will be seen by the ThreadSMR thread
>      >      iterator but will have a NULL java.lang.Thread object.
>      >
>      >      We special-case attaching threads in a number of places in the VM and I
>      >      think we should be explicitly doing something here to filter out
>      >      attaching threads, rather than just being tolerant of a NULL j.l.Thread
>      >      object. Specifically in ThreadsSMRSupport::add_thread:
>      >
>      >      if (ThreadTable::is_initialized() && !thread->is_attaching_via_jni()) {
>      >         jlong tid = java_lang_Thread::thread_id(thread->threadObj());
>      >         ThreadTable::add_thread(tid, thread);
>      >      }
>      >
>      >      Note that in ThreadsSMRSupport::remove_thread we can use the same guard,
>      >      which covers the case the JNI attach encountered an error trying to
>      >      create the j.l.Thread object.
>      >
>      >      >> src/hotspot/share/services/threadTable.cpp
>      >      >> 71     static uintx get_hash(Value const& value, bool* is_dead) {
>      >      >
>      >      >> The is_dead parameter still bothers me here. I can't make enough sense
>      >      >> out of the template code in ConcurrentHashtable to see why we have to
>      >      >> have it, but I'm concerned that its very existence means we perhaps
>      >      >> should not be trying to extend CHT in this context. ??
>      >      >
>      >      > My understanding is that is_dead parameter provides a mechanism for
>      >      > ConcurrentHashtable to remove stale entries that were not explicitly
>      >      > removed by calling  ConcurrentHashTable::remove() method.
>      >      > I think that just because in our case we don't use this mechanism doesn't
>      >      > mean we should not use ConcurrentHashTable.
>      >
>      >      Can you confirm that this usage is okay with Robbin Ehn please. He's
>      >      back from vacation this week.
>      >
>      >      >> I would still want to see what impact this has on thread
>      >      >> startup cost, both with and without the table being initialized.
>      >      >
>      >      > I run a test that initializes the table by calling ThreadMXBean.get getThreadInfo(),
>      >      > starts some threads as a worm-up, and then creates and starts 100,000 threads
>      >      > (each thread just sleeps for 100 ms). In case when the thread table is enabled
>      >      > 100,000 threads are created and started  for about 15200 ms. If the thread table
>      >      > is off the test takes about 14800 ms. Based on this information the enabled
>      >      > thread table makes the thread startup about 2.7% slower.
>      >
>      >      That doesn't sound very good. I think we may need to Claes involved to
>      >      help investigate overall performance impact here.
>      >
>      >      > Webrev: https://cr.openjdk.java.net/~dtitov/8185005/webrev.04/
>      >      > Bug: https://bugs.openjdk.java.net/browse/JDK-8185005
>      >
>      >      No further code comments.
>      >
>      >      I didn't look at the test in detail.
>      >
>      >      Thanks,
>      >      David
>      >
>      >      > Thanks!
>      >      > --Daniil
>      >      >
>      >      >
>      >      > On 7/29/19, 12:53 AM, "David Holmes" <[hidden email]> wrote:
>      >      >
>      >      >      Hi Daniil,
>      >      >
>      >      >      Overall I think this is a reasonable approach but I would still like to
>      >      >      see some performance and footprint numbers, both to verify it fixes the
>      >      >      problem reported, and that we are not getting penalized elsewhere.
>      >      >
>      >      >      On 25/07/2019 3:21 am, Daniil Titov wrote:
>      >      >      > Hi David, Daniel, and Serguei,
>      >      >      >
>      >      >      > Please review the new version of the fix, that makes the thread table initialization on demand and
>      >      >      > moves it inside ThreadsList::find_JavaThread_from_java_tid(). At the creation time the thread table
>      >      >      >   is initialized with the threads from the current thread list. We don't want to hold Threads_lock
>      >      >      > inside find_JavaThread_from_java_tid(),  thus new threads still could be created  while the thread
>      >      >      > table is being initialized . Such threads will be found by the linear search and added to the thread table
>      >      >      > later, in ThreadsList::find_JavaThread_from_java_tid().
>      >      >
>      >      >      The initialization allows the created but unpopulated, or partially
>      >      >      populated, table to be seen by other threads - is that your intention?
>      >      >      It seems it should be okay as the other threads will then race with the
>      >      >      initializing thread to add specific entries, and this is a concurrent
>      >      >      map so that should be functionally correct. But if so then I think you
>      >      >      can also reduce the scope of the ThreadTableCreate_lock so that it
>      >      >      covers creation of the table only, not the initial population of the table.
>      >      >
>      >      >      I like the approach of only initializing the table when needed and using
>      >      >      that to control when the add/remove-thread code needs to update the
>      >      >      table. But I would still want to see what impact this has on thread
>      >      >      startup cost, both with and without the table being initialized.
>      >      >
>      >      >      > The change also includes additional optimization for some callers of find_JavaThread_from_java_tid()
>      >      >      > as Daniel suggested.
>      >      >
>      >      >      Not sure it's best to combine these, but if they are limited to the
>      >      >      changes in management.cpp only then that may be okay. It helps to be
>      >      >      able to focus on the table related changes without being distracted by
>      >      >      other optimizations.
>      >      >
>      >      >      > That is correct that ResolvedMethodTable was used as a blueprint for the thread table, however, I tried
>      >      >      > to strip it of the all functionality that is not required in the thread table case.
>      >      >
>      >      >      The revised version seems better in that regard. But I still have a
>      >      >      concern, see below.
>      >      >
>      >      >      > We need to have the thread table resizable and allow it to grow as the number of threads increases to avoid
>      >      >      > reserving excessive memory a-priori or deteriorating lookup times. The ServiceThread is responsible for
>      >      >      > growing the thread table when required.
>      >      >
>      >      >      Yes but why? Why can't this table be grown on demand by the thread that
>      >      >      is doing the addition? For other tables we may have to delegate to the
>      >      >      service thread because the current thread cannot perform the action, or
>      >      >      it doesn't want to perform it at the time the need for the resize is
>      >      >      detected (e.g. its detected at a safepoint and you want the resize to
>      >      >      happen later outside the safepoint). It's not apparent to me that such
>      >      >      restrictions apply here.
>      >      >
>      >      >      > There is no ConcurrentHashTable available in Java 8 and for backporting this fix to Java 8 another implementation
>      >      >      > of the hash table, probably originally suggested in the patch attached to the JBS issue, should be used.  It will make
>      >      >      > the backporting more complicated,  however, adding a new Implementation of the hash table in Java 14 while it
>      >      >      > already has ConcurrentHashTable doesn't seem  reasonable for me.
>      >      >
>      >      >      Ok.
>      >      >
>      >      >      > Webrev: http://cr.openjdk.java.net/~dtitov/8185005/webrev.03
>      >      >
>      >      >      Some specific code comments:
>      >      >
>      >      >      src/hotspot/share/runtime/mutexLocker.cpp
>      >      >
>      >      >      +   def(ThreadTableCreate_lock       , PaddedMutex  , special,
>      >      >      false, Monitor::_safepoint_check_never);
>      >      >
>      >      >      I think this needs to be a _safepoint_check_always lock. The table will
>      >      >      be created by regular JavaThreads and they should (nearly) always be
>      >      >      checking for safepoints if they are going to block acquiring the lock.
>      >      >      And it isn't at all obvious that the thread doing the creation can't go
>      >      >      to a safepoint whilst this lock is held.
>      >      >
>      >      >      ---
>      >      >
>      >      >      src/hotspot/share/runtime/threadSMR.cpp
>      >      >
>      >      >      Nit:
>      >      >
>      >      >        618       JavaThread* thread = thread_at(i);
>      >      >
>      >      >      you could reuse the new java_thread local you introduced at line 613 and
>      >      >      just rename that "new" variable to "thread" so you don't have to change
>      >      >      all other uses.
>      >      >
>      >      >      628   } else if (java_thread != NULL && ...
>      >      >
>      >      >      You don't need to check != NULL here as you only get here when
>      >      >      java_thread is not NULL.
>      >      >
>      >      >        755     jlong tid = SharedRuntime::get_java_tid(thread);
>      >      >        926     jlong tid = SharedRuntime::get_java_tid(thread);
>      >      >
>      >      >      I think it cleaner/better to just use
>      >      >
>      >      >      jlong tid = java_lang_Thread::thread_id(thread->threadObj());
>      >      >
>      >      >      as we know thread is not NULL, it is a JavaThread and it has to have a
>      >      >      non-null threadObj.
>      >      >
>      >      >      ---
>      >      >
>      >      >      src/hotspot/share/services/management.cpp
>      >      >
>      >      >      1323         if (THREAD->is_Java_thread()) {
>      >      >      1324           JavaThread* current_thread = (JavaThread*)THREAD;
>      >      >
>      >      >      These calls can only be made on a JavaThread so this be simplified to
>      >      >      remove the is_Java_thread() call. Similarly in other places.
>      >      >
>      >      >      ---
>      >      >
>      >      >      src/hotspot/share/services/threadTable.cpp
>      >      >
>      >      >         55 class ThreadTableEntry : public CHeapObj<mtInternal> {
>      >      >         56   private:
>      >      >         57     jlong _tid;
>      >      >
>      >      >      I believe hotspot style is to not indent the access modifiers in C++
>      >      >      class declarations, so the above would just be:
>      >      >
>      >      >         55 class ThreadTableEntry : public CHeapObj<mtInternal> {
>      >      >         56 private:
>      >      >         57   jlong _tid;
>      >      >
>      >      >      etc.
>      >      >
>      >      >        60     ThreadTableEntry(jlong tid, JavaThread* java_thread) :
>      >      >        61     _tid(tid),_java_thread(java_thread) {}
>      >      >
>      >      >      line 61 should be indented as it continues line 60.
>      >      >
>      >      >         67 class ThreadTableConfig : public AllStatic {
>      >      >         ...
>      >      >         71     static uintx get_hash(Value const& value, bool* is_dead) {
>      >      >
>      >      >      The is_dead parameter still bothers me here. I can't make enough sense
>      >      >      out of the template code in ConcurrentHashtable to see why we have to
>      >      >      have it, but I'm concerned that its very existence means we perhaps
>      >      >      should not be trying to extend CHT in this context. ??
>      >      >
>      >      >        115   size_t start_size_log = size_log > DefaultThreadTableSizeLog
>      >      >        116   ? size_log : DefaultThreadTableSizeLog;
>      >      >
>      >      >      line 116 should be indented, though in this case I think a better layout
>      >      >      would be:
>      >      >
>      >      >        115   size_t start_size_log =
>      >      >        116       size_log > DefaultThreadTableSizeLog ? size_log :
>      >      >      DefaultThreadTableSizeLog;
>      >      >
>      >      >        131 double ThreadTable::get_load_factor() {
>      >      >        132   return (double)_items_count/_current_size;
>      >      >        133 }
>      >      >
>      >      >      Not sure that is doing what you want/expect. It will perform integer
>      >      >      division and then cast that whole integer to a double. If you want
>      >      >      double arithmetic you need:
>      >      >
>      >      >      return ((double)_items_count)/_current_size;
>      >      >
>      >      >      180     jlong          _tid;
>      >      >      181     uintx         _hash;
>      >      >
>      >      >      Nit: no need for all those spaces before the variable name.
>      >      >
>      >      >        183     ThreadTableLookup(jlong tid)
>      >      >        184     : _tid(tid), _hash(primitive_hash(tid)) {}
>      >      >
>      >      >      line 184 should be indented.
>      >      >
>      >      >      201     ThreadGet():_return(NULL) {}
>      >      >
>      >      >      Nit: need space after :
>      >      >
>      >      >        211    assert(_is_initialized, "Thread table is not initialized");
>      >      >        212   _has_work = false;
>      >      >
>      >      >      line 211 is indented one space too far.
>      >      >
>      >      >      229     ThreadTableEntry* entry = new ThreadTableEntry(tid,java_thread);
>      >      >
>      >      >      Nit: need space after ,
>      >      >
>      >      >      252   return _local_table->remove(thread,lookup);
>      >      >
>      >      >      Nit: need space after ,
>      >      >
>      >      >      Thanks,
>      >      >      David
>      >      >      ------
>      >      >
>      >      >      > Bug: https://bugs.openjdk.java.net/browse/JDK-8185005
>      >      >      >
>      >      >      > Thanks!
>      >      >      > --Daniil
>      >      >      >
>      >      >      >
>      >      >      > On 7/8/19, 3:24 PM, "Daniel D. Daugherty" <[hidden email]> wrote:
>      >      >      >
>      >      >      >      On 6/29/19 12:06 PM, Daniil Titov wrote:
>      >      >      >      > Hi Serguei and David,
>      >      >      >      >
>      >      >      >      > Serguei is right, ThreadTable::find_thread(java_tid) cannot  return a JavaThread with an unmatched java_tid.
>      >      >      >      >
>      >      >      >      > Please find a new version of the fix that includes the changes Serguei suggested.
>      >      >      >      >
>      >      >      >      > Regarding the concern about the maintaining the thread table when it may never even be queried, one of
>      >      >      >      > the options could be to add ThreadTable ::isEnabled flag, set it to "false" by default, and wrap the calls to the thread table
>      >      >      >      > in ThreadsSMRSupport add_thread() and remove_thread() methods to check this flag.
>      >      >      >      >
>      >      >      >      > When ThreadsList::find_JavaThread_from_java_tid() is called for the first time it could check if ThreadTable ::isEnabled
>      >      >      >      > Is on and if not then set it on and populate the thread table with all existing threads from the thread list.
>      >      >      >
>      >      >      >      I have the same concerns as David H. about this new ThreadTable.
>      >      >      >      ThreadsList::find_JavaThread_from_java_tid() is only called from code
>      >      >      >      in src/hotspot/share/services/management.cpp so I think that table
>      >      >      >      needs to enabled and populated only if it is going to be used.
>      >      >      >
>      >      >      >      I've taken a look at the webrev below and I see that David has
>      >      >      >      followed up with additional comments. Before I do a crawl through
>      >      >      >      code review for this, I would like to see the ThreadTable stuff
>      >      >      >      made optional and David's other comments addressed.
>      >      >      >
>      >      >      >      Another possible optimization is for callers of
>      >      >      >      find_JavaThread_from_java_tid() to save the calling thread's
>      >      >      >      tid value before they loop and if the current tid == saved_tid
>      >      >      >      then use the current JavaThread* instead of calling
>      >      >      >      find_JavaThread_from_java_tid() to get the JavaThread*.
>      >      >      >
>      >      >      >      Dan
>      >      >      >
>      >      >      >      >
>      >      >      >      > Webrev: https://cr.openjdk.java.net/~dtitov/8185005/webrev.02/
>      >      >      >      > Bug: https://bugs.openjdk.java.net/browse/JDK-8185005
>      >      >      >      >
>      >      >      >      > Thanks!
>      >      >      >      > --Daniil
>      >      >      >      >
>      >      >      >      > From: <[hidden email]>
>      >      >      >      > Organization: Oracle Corporation
>      >      >      >      > Date: Friday, June 28, 2019 at 7:56 PM
>      >      >      >      > To: Daniil Titov <[hidden email]>, OpenJDK Serviceability <[hidden email]>, "[hidden email]" <[hidden email]>, "[hidden email]" <[hidden email]>
>      >      >      >      > Subject: Re: RFR: 8185005: Improve performance of ThreadMXBean.getThreadInfo(long ids[], int maxDepth)
>      >      >      >      >
>      >      >      >      > Hi Daniil,
>      >      >      >      >
>      >      >      >      > I have several quick comments.
>      >      >      >      >
>      >      >      >      > The indent in the hotspot c/c++ files has to be 2, not 4.
>      >      >      >      >
>      >      >      >      > https://cr.openjdk.java.net/~dtitov/8185005/webrev.01/src/hotspot/share/runtime/threadSMR.cpp.frames.html
>      >      >      >      > 614 JavaThread* ThreadsList::find_JavaThread_from_java_tid(jlong java_tid) const {
>      >      >      >      >   615     JavaThread* java_thread = ThreadTable::find_thread(java_tid);
>      >      >      >      >   616     if (java_thread == NULL && java_tid == PMIMORDIAL_JAVA_TID) {
>      >      >      >      >   617         // ThreadsSMRSupport::add_thread() is not called for the primordial
>      >      >      >      >   618         // thread. Thus, we find this thread with a linear search and add it
>      >      >      >      >   619         // to the thread table.
>      >      >      >      >   620         for (uint i = 0; i < length(); i++) {
>      >      >      >      >   621             JavaThread* thread = thread_at(i);
>      >      >      >      >   622             if (is_valid_java_thread(java_tid,thread)) {
>      >      >      >      >   623                 ThreadTable::add_thread(java_tid, thread);
>      >      >      >      >   624                 return thread;
>      >      >      >      >   625             }
>      >      >      >      >   626         }
>      >      >      >      >   627     } else if (java_thread != NULL && is_valid_java_thread(java_tid, java_thread)) {
>      >      >      >      >   628         return java_thread;
>      >      >      >      >   629     }
>      >      >      >      >   630     return NULL;
>      >      >      >      >   631 }
>      >      >      >      >   632 bool ThreadsList::is_valid_java_thread(jlong java_tid, JavaThread* java_thread) {
>      >      >      >      >   633     oop tobj = java_thread->threadObj();
>      >      >      >      >   634     // Ignore the thread if it hasn't run yet, has exited
>      >      >      >      >   635     // or is starting to exit.
>      >      >      >      >   636     return (tobj != NULL && !java_thread->is_exiting() &&
>      >      >      >      >   637             java_tid == java_lang_Thread::thread_id(tobj));
>      >      >      >      >   638 }
>      >      >      >      >
>      >      >      >      >   615     JavaThread* java_thread = ThreadTable::find_thread(java_tid);
>      >      >      >      >
>      >      >      >      >    I'd suggest to rename find_thread() to find_thread_by_tid().
>      >      >      >      >
>      >      >      >      > A space is missed after the comma:
>      >      >      >      >    622 if (is_valid_java_thread(java_tid,thread)) {
>      >      >      >      >
>      >      >      >      > An empty line is needed before L632.
>      >      >      >      >
>      >      >      >      > The name 'is_valid_java_thread' looks wrong (or confusing) to me.
>      >      >      >      > Something like 'is_alive_java_thread_with_tid()' would be better.
>      >      >      >      > It'd better to list parameters in the opposite order.
>      >      >      >      >
>      >      >      >      > The call to is_valid_java_thread() is confusing:
>      >      >      >      >     627 } else if (java_thread != NULL && is_valid_java_thread(java_tid, java_thread)) {
>      >      >      >      >
>      >      >      >      > Why would the call ThreadTable::find_thread(java_tid) return a JavaThread with an unmatched java_tid?
>      >      >      >      >
>      >      >      >      >
>      >      >      >      > Thanks,
>      >      >      >      > Serguei
>      >      >      >      >
>      >      >      >      > On 6/28/19, 9:40 PM, "David Holmes" <[hidden email]> wrote:
>      >      >      >      >
>      >      >      >      >      Hi Daniil,
>      >      >      >      >
>      >      >      >      >      The definition and use of this hashtable (yet another hashtable
>      >      >      >      >      implementation!) will need careful examination. We have to be concerned
>      >      >      >      >      about the cost of maintaining it when it may never even be queried. You
>      >      >      >      >      would need to look at footprint cost and performance impact.
>      >      >      >      >
>      >      >      >      >      Unfortunately I'm just about to board a plane and will be out for the
>      >      >      >      >      next few days. I will try to look at this asap next week, but we will
>      >      >      >      >      need a lot more data on it.
>      >      >      >      >
>      >      >      >      >      Thanks,
>      >      >      >      >      David
>      >      >      >      >
>      >      >      >      > On 6/28/19 3:31 PM, Daniil Titov wrote:
>      >      >      >      > Please review the change that improves performance of ThreadMXBean MXBean methods returning the
>      >      >      >      > information for specific threads. The change introduces the thread table that uses ConcurrentHashTable
>      >      >      >      > to store one-to-one the mapping between the thread ids and JavaThread objects and replaces the linear
>      >      >      >      > search over the thread list in ThreadsList::find_JavaThread_from_java_tid(jlong tid) method with the lookup
>      >      >      >      > in the thread table.
>      >      >      >      >
>      >      >      >      > Testing: Mach5 tier1,tier2 and tier3 tests successfully passed.
>      >      >      >      >
>      >      >      >      > Webrev: https://cr.openjdk.java.net/~dtitov/8185005/webrev.01/
>      >      >      >      > Bug: https://bugs.openjdk.java.net/browse/JDK-8185005
>      >      >      >      >
>      >      >      >      > Thanks!
>      >      >      >      >
>      >      >      >      > Best regards,
>      >      >      >      > Daniil
>      >      >      >      >
>      >      >      >      >
>      >      >      >      >
>      >      >      >      >
>      >      >      >      >
>      >      >      >      >
>      >      >      >      >
>      >      >      >
>      >      >      >
>      >      >      >
>      >      >      >
>      >      >
>      >      >
>      >      >
>      >
>      >
>      >
>      
>
>
Reply | Threaded
Open this post in threaded view
|

Re: 8185005: Improve performance of ThreadMXBean.getThreadInfo(long ids[], int maxDepth)

Daniel D. Daugherty
So based on the last three message on this thread:

webrev.05 is withdrawn for the moment
webrev.04 is the current webrev, but needs to have some startup time
           issues resolved before moving forward.

I'm going to hold off on re-reviewing anything at the moment until
the dust settles...

Dan


On 8/12/19 7:34 PM, David Holmes wrote:

> Hi Daniil,
>
> On 13/08/2019 9:24 am, Daniil Titov wrote:
>> Hi Robbin,
>>
>> Thank you very much for reviewing this version of the fix! Based on
>> your findings
>> it seems as it makes sense to make a step back and  continue with the
>> approach we took before in the previous version of the webrev
>> (webrev.04),
>> and get more information about the impact on the startup time it has.
>> I will
>> consult with Claus regarding this and then share the findings.
>
> That seems a good approach to me. It wasn't at all clear to me that
> the latest proposed approach would actually solve the original problem
> in a satisfactory way - it would depend on how constant the set of
> threads being queried was.
>
> There is no perfect solution here as any fix to the reported problem
> incurs overhead elsewhere. Even evaluating the merits of the different
> trade-offs is hard to do - we could end up with a compromise solution
> that fails to satisfy anyone.
>
> David
> -----
>
>> Thanks again,
>> --Daniil
>>
>>
>>
>>
>>
>>
>> On 8/12/19, 5:22 AM, "Robbin Ehn" <[hidden email]> wrote:
>>
>>      Hi Daniil,
>>           I took a new deeper dive into this.
>>           This line seems to have some issues:
>>           if (ThreadTable::is_initialized() &&
>> thread->in_thread_table() &&
>>      !thread->is_attaching_via_jni()) {
>>           If you create new threads which attaches and then dies, the
>> table will just keep
>>      growing. So you must remove them also ?
>>           Secondly you should not use volatile semantics for
>> _in_thread_table.
>>      The load in the if-statement can be reordered with _is_initialized.
>>      Which could lead to a leak, rogue pointer in the table.
>>           So both "static volatile bool _is_initialized;" and
>> "volatile bool
>>      _in_thread_table; "
>>      should be stored with store_release and loaded with load_acquire.
>>           Unfortunately it looks like there still would be races if
>>      ThreadTable::add_thread e.g. context switch at:
>>           if (_local_table->insert(thread, lookup, entry)) {
>>      // HERE
>>         java_thread->set_in_thread_table(true);
>>           *Remove side can pass the if-statement without removing.
>>           Since this thread also maybe exiting at any moment, e.g.
>> context switch:
>>                  if (tobj != NULL && !thread->is_exiting() &&
>>                 java_tid == java_lang_Thread::thread_id(tobj)) {
>>          // HERE
>>               ThreadTable::add_thread(java_tid, thread);
>>           *Add side can add a thread that is exiting.
>>           Mixing in a third thread looking up a random tid and
>> getting a JavaThread*, it
>>      must validate it against it's ThreadsList. Making the hashtable
>> useless.
>>           So I think the only one adding and removing should be the
>> thread itself.
>>      1:Add to ThreadsList
>>      2:Add to ThreadTable
>>      3:Remove from ThreadTable
>>      4:Remove ThreadsList
>>           Between 1-2 and 3-4 the thread would be looked-up via
>> linear scan.
>>      I don't see an easy way around the start-up issue with this.
>>           Maybe have the cache in Java.
>>      Pass in the thread obj into a
>>      java_sun_management_ThreadImpl_getThreadTotalCpuTime3 instead,
>>      thus skipping any look-ups in native.
>>           Thanks, Robbin
>>                On 8/12/19 5:49 AM, Daniil Titov wrote:
>>      > Hi David, Robbin, Daniel, and Serguei,
>>      >
>>      > Please review a new version of the fix.
>>      >
>>      > As David suggested I created a separated Jira issue [1] to
>> cover  additional optimization for
>>      > some callers of find_JavaThread_from_java_tid() and this
>> version of the fix no longer includes
>>      > changes in management.cpp ( and the test related with these
>> changes).
>>      >
>>      > Regarding the impact the previous version of the fix had on
>> the thread startup time at heavy load (e.g.
>>      > when 5000 threads are created and destroyed every second) I
>> tried a different approach that makes
>>      > calls to ThreadTable::add_thread  and
>> ThreadTable::remove_thread  asynchronous and offloads the
>>      > work for actual modifications of the thread table to a
>> periodic task that runs every 5 seconds. With the
>>      > same  stress test scenario (the  test does some warm-up and
>> then measures the time it takes to create
>>      > and start 100,000 threads; every  thread just sleeps for 100
>> ms) the impact on the thread startup time
>>      > was reduced to 1.2% ( from 2.7%).
>>      >
>>      > The cause of this impact in this stress test scenario is that
>> as soon as the thread table is initialized,
>>      > an additional work to insert  and delete entries in the thread
>> table should be performed, even if
>>      > com.sun.management.ThreadMXBean methods are no longer called.
>> For example, In the stress test
>>      > mentioned above, every second about 5000 entries had to be
>> inserted in the table and then deleted.
>>      >
>>      > That doesn't look right and the new version of the fix uses
>> the different approach: the thread is added to
>>      > the thread table only when this thread is requested by
>> com.sun.management.ThreadMXBean bean. Every
>>      > time when find_JavaThread_from_java_tid() is called for a new
>> tid, the thread  is found by the iterating over
>>      > the thread list and added to the thread table. All consequent
>> calls to find_JavaThread_from_java_tid() for
>>      > the same tid returns the thread from the thread table.
>>      >
>>      > Running stress test for the cases when the thread table is
>> enabled and not showed no difference in the
>>      > average thread startup times.
>>      >
>>      > [1] : https://bugs.openjdk.java.net/browse/JDK-8229391
>>      >
>>      > Bug: https://bugs.openjdk.java.net/browse/JDK-8185005
>>      > Webrev: https://cr.openjdk.java.net/~dtitov/8185005/webrev.05/
>>      >
>>      > Thanks,
>>      > Daniil
>>      >
>>      > On 8/4/19, 7:54 PM, "David Holmes" <[hidden email]>
>> wrote:
>>      >
>>      >      Hi Daniil,
>>      >
>>      >      On 3/08/2019 8:16 am, Daniil Titov wrote:
>>      >      > Hi David,
>>      >      >
>>      >      > Thank you for your detailed review. Please review a new
>> version of the fix that includes
>>      >      > the changes you suggested:
>>      >      > - ThreadTableCreate_lock scope is reduced to cover the
>> creation of the table only;
>>      >      > - ThreadTableCreate_lock is made _safepoint_check_always;
>>      >
>>      >      Okay.
>>      >
>>      >      > - ServiceThread is no longer responsible for the
>> resizing of the thread table, instead,
>>      >      >    the thread table is changed to grow on demand by the
>> thread that is doing the addition;
>>      >
>>      >      Okay - I'm happy to get the serviceThread out of the
>> picture here.
>>      >
>>      >      > - fixed nits and formatting issues.
>>      >
>>      >      Okay.
>>      >
>>      >      >>> The change also includes additional optimization for
>> some callers of find_JavaThread_from_java_tid()
>>      >      >>>   as Daniel suggested.
>>      >      >> Not sure it's best to combine these, but if they are
>> limited to the
>>      >      >> changes in management.cpp only then that may be okay.
>>      >      >
>>      >      > The additional optimization for some callers of
>> find_JavaThread_from_java_tid() is
>>      >      > limited to management.cpp (plus a new test) so I left
>> them in the webrev  but
>>      >      > I also could move it in the separate issue if required.
>>      >
>>      >      I'd prefer this part of be separated out, but won't
>> insist. Let's see if
>>      >      Dan or Serguei have a strong opinion.
>>      >
>>      >      >    > src/hotspot/share/runtime/threadSMR.cpp
>>      >      >    >755     jlong tid =
>> SharedRuntime::get_java_tid(thread);
>>      >      >    > 926     jlong tid =
>> SharedRuntime::get_java_tid(thread);
>>      >      >   >  I think it cleaner/better to just use
>>      >      >   > jlong tid =
>> java_lang_Thread::thread_id(thread->threadObj());
>>      >      >   > as we know thread is not NULL, it is a JavaThread
>> and it has to have a
>>      >      >   > non-null threadObj.
>>      >      >
>>      >      > I had to leave this code unchanged since it turned out
>> the threadObj is null
>>      >      > when VM is destroyed:
>>      >      >
>>      >      > V  [libjvm.so+0xe165d7] oopDesc::long_field(int)
>> const+0x67
>>      >      > V  [libjvm.so+0x16e06c6]
>> ThreadsSMRSupport::add_thread(JavaThread*)+0x116
>>      >      > V  [libjvm.so+0x16d1302] Threads::add(JavaThread*,
>> bool)+0x82
>>      >      > V  [libjvm.so+0xef8369]
>> attach_current_thread.part.197+0xc9
>>      >      > V  [libjvm.so+0xec136c] jni_DestroyJavaVM+0x6c
>>      >      > C  [libjli.so+0x4333]  JavaMain+0x2c3
>>      >      > C  [libjli.so+0x8159]  ThreadJavaMain+0x9
>>      >
>>      >      This is actually nothing to do with the VM being
>> destroyed, but is an
>>      >      issue with JNI_AttachCurrentThread and its interaction
>> with the
>>      >      ThreadSMR iterators. The attach process is:
>>      >      - create JavaThread
>>      >      - mark as "is attaching via jni"
>>      >      - add to ThreadsList
>>      >      - create java.lang.Thread object (you can only execute
>> Java code after
>>      >      you are attached)
>>      >      - mark as "attach completed"
>>      >
>>      >      So while a thread "is attaching" it will be seen by the
>> ThreadSMR thread
>>      >      iterator but will have a NULL java.lang.Thread object.
>>      >
>>      >      We special-case attaching threads in a number of places
>> in the VM and I
>>      >      think we should be explicitly doing something here to
>> filter out
>>      >      attaching threads, rather than just being tolerant of a
>> NULL j.l.Thread
>>      >      object. Specifically in ThreadsSMRSupport::add_thread:
>>      >
>>      >      if (ThreadTable::is_initialized() &&
>> !thread->is_attaching_via_jni()) {
>>      >         jlong tid =
>> java_lang_Thread::thread_id(thread->threadObj());
>>      >         ThreadTable::add_thread(tid, thread);
>>      >      }
>>      >
>>      >      Note that in ThreadsSMRSupport::remove_thread we can use
>> the same guard,
>>      >      which covers the case the JNI attach encountered an error
>> trying to
>>      >      create the j.l.Thread object.
>>      >
>>      >      >> src/hotspot/share/services/threadTable.cpp
>>      >      >> 71     static uintx get_hash(Value const& value, bool*
>> is_dead) {
>>      >      >
>>      >      >> The is_dead parameter still bothers me here. I can't
>> make enough sense
>>      >      >> out of the template code in ConcurrentHashtable to see
>> why we have to
>>      >      >> have it, but I'm concerned that its very existence
>> means we perhaps
>>      >      >> should not be trying to extend CHT in this context. ??
>>      >      >
>>      >      > My understanding is that is_dead parameter provides a
>> mechanism for
>>      >      > ConcurrentHashtable to remove stale entries that were
>> not explicitly
>>      >      > removed by calling ConcurrentHashTable::remove() method.
>>      >      > I think that just because in our case we don't use this
>> mechanism doesn't
>>      >      > mean we should not use ConcurrentHashTable.
>>      >
>>      >      Can you confirm that this usage is okay with Robbin Ehn
>> please. He's
>>      >      back from vacation this week.
>>      >
>>      >      >> I would still want to see what impact this has on thread
>>      >      >> startup cost, both with and without the table being
>> initialized.
>>      >      >
>>      >      > I run a test that initializes the table by calling
>> ThreadMXBean.get getThreadInfo(),
>>      >      > starts some threads as a worm-up, and then creates and
>> starts 100,000 threads
>>      >      > (each thread just sleeps for 100 ms). In case when the
>> thread table is enabled
>>      >      > 100,000 threads are created and started  for about
>> 15200 ms. If the thread table
>>      >      > is off the test takes about 14800 ms. Based on this
>> information the enabled
>>      >      > thread table makes the thread startup about 2.7% slower.
>>      >
>>      >      That doesn't sound very good. I think we may need to
>> Claes involved to
>>      >      help investigate overall performance impact here.
>>      >
>>      >      > Webrev:
>> https://cr.openjdk.java.net/~dtitov/8185005/webrev.04/
>>      >      > Bug: https://bugs.openjdk.java.net/browse/JDK-8185005
>>      >
>>      >      No further code comments.
>>      >
>>      >      I didn't look at the test in detail.
>>      >
>>      >      Thanks,
>>      >      David
>>      >
>>      >      > Thanks!
>>      >      > --Daniil
>>      >      >
>>      >      >
>>      >      > On 7/29/19, 12:53 AM, "David Holmes"
>> <[hidden email]> wrote:
>>      >      >
>>      >      >      Hi Daniil,
>>      >      >
>>      >      >      Overall I think this is a reasonable approach but
>> I would still like to
>>      >      >      see some performance and footprint numbers, both
>> to verify it fixes the
>>      >      >      problem reported, and that we are not getting
>> penalized elsewhere.
>>      >      >
>>      >      >      On 25/07/2019 3:21 am, Daniil Titov wrote:
>>      >      >      > Hi David, Daniel, and Serguei,
>>      >      >      >
>>      >      >      > Please review the new version of the fix, that
>> makes the thread table initialization on demand and
>>      >      >      > moves it inside
>> ThreadsList::find_JavaThread_from_java_tid(). At the creation time
>> the thread table
>>      >      >      >   is initialized with the threads from the
>> current thread list. We don't want to hold Threads_lock
>>      >      >      > inside find_JavaThread_from_java_tid(),  thus
>> new threads still could be created  while the thread
>>      >      >      > table is being initialized . Such threads will
>> be found by the linear search and added to the thread table
>>      >      >      > later, in
>> ThreadsList::find_JavaThread_from_java_tid().
>>      >      >
>>      >      >      The initialization allows the created but
>> unpopulated, or partially
>>      >      >      populated, table to be seen by other threads - is
>> that your intention?
>>      >      >      It seems it should be okay as the other threads
>> will then race with the
>>      >      >      initializing thread to add specific entries, and
>> this is a concurrent
>>      >      >      map so that should be functionally correct. But if
>> so then I think you
>>      >      >      can also reduce the scope of the
>> ThreadTableCreate_lock so that it
>>      >      >      covers creation of the table only, not the initial
>> population of the table.
>>      >      >
>>      >      >      I like the approach of only initializing the table
>> when needed and using
>>      >      >      that to control when the add/remove-thread code
>> needs to update the
>>      >      >      table. But I would still want to see what impact
>> this has on thread
>>      >      >      startup cost, both with and without the table
>> being initialized.
>>      >      >
>>      >      >      > The change also includes additional optimization
>> for some callers of find_JavaThread_from_java_tid()
>>      >      >      > as Daniel suggested.
>>      >      >
>>      >      >      Not sure it's best to combine these, but if they
>> are limited to the
>>      >      >      changes in management.cpp only then that may be
>> okay. It helps to be
>>      >      >      able to focus on the table related changes without
>> being distracted by
>>      >      >      other optimizations.
>>      >      >
>>      >      >      > That is correct that ResolvedMethodTable was
>> used as a blueprint for the thread table, however, I tried
>>      >      >      > to strip it of the all functionality that is not
>> required in the thread table case.
>>      >      >
>>      >      >      The revised version seems better in that regard.
>> But I still have a
>>      >      >      concern, see below.
>>      >      >
>>      >      >      > We need to have the thread table resizable and
>> allow it to grow as the number of threads increases to avoid
>>      >      >      > reserving excessive memory a-priori or
>> deteriorating lookup times. The ServiceThread is responsible for
>>      >      >      > growing the thread table when required.
>>      >      >
>>      >      >      Yes but why? Why can't this table be grown on
>> demand by the thread that
>>      >      >      is doing the addition? For other tables we may
>> have to delegate to the
>>      >      >      service thread because the current thread cannot
>> perform the action, or
>>      >      >      it doesn't want to perform it at the time the need
>> for the resize is
>>      >      >      detected (e.g. its detected at a safepoint and you
>> want the resize to
>>      >      >      happen later outside the safepoint). It's not
>> apparent to me that such
>>      >      >      restrictions apply here.
>>      >      >
>>      >      >      > There is no ConcurrentHashTable available in
>> Java 8 and for backporting this fix to Java 8 another implementation
>>      >      >      > of the hash table, probably originally suggested
>> in the patch attached to the JBS issue, should be used.  It will make
>>      >      >      > the backporting more complicated, however,
>> adding a new Implementation of the hash table in Java 14 while it
>>      >      >      > already has ConcurrentHashTable doesn't seem 
>> reasonable for me.
>>      >      >
>>      >      >      Ok.
>>      >      >
>>      >      >      > Webrev:
>> http://cr.openjdk.java.net/~dtitov/8185005/webrev.03
>>      >      >
>>      >      >      Some specific code comments:
>>      >      >
>>      >      > src/hotspot/share/runtime/mutexLocker.cpp
>>      >      >
>>      >      >      +   def(ThreadTableCreate_lock       ,
>> PaddedMutex  , special,
>>      >      >      false, Monitor::_safepoint_check_never);
>>      >      >
>>      >      >      I think this needs to be a _safepoint_check_always
>> lock. The table will
>>      >      >      be created by regular JavaThreads and they should
>> (nearly) always be
>>      >      >      checking for safepoints if they are going to block
>> acquiring the lock.
>>      >      >      And it isn't at all obvious that the thread doing
>> the creation can't go
>>      >      >      to a safepoint whilst this lock is held.
>>      >      >
>>      >      >      ---
>>      >      >
>>      >      >      src/hotspot/share/runtime/threadSMR.cpp
>>      >      >
>>      >      >      Nit:
>>      >      >
>>      >      >        618       JavaThread* thread = thread_at(i);
>>      >      >
>>      >      >      you could reuse the new java_thread local you
>> introduced at line 613 and
>>      >      >      just rename that "new" variable to "thread" so you
>> don't have to change
>>      >      >      all other uses.
>>      >      >
>>      >      >      628   } else if (java_thread != NULL && ...
>>      >      >
>>      >      >      You don't need to check != NULL here as you only
>> get here when
>>      >      >      java_thread is not NULL.
>>      >      >
>>      >      >        755     jlong tid =
>> SharedRuntime::get_java_tid(thread);
>>      >      >        926     jlong tid =
>> SharedRuntime::get_java_tid(thread);
>>      >      >
>>      >      >      I think it cleaner/better to just use
>>      >      >
>>      >      >      jlong tid =
>> java_lang_Thread::thread_id(thread->threadObj());
>>      >      >
>>      >      >      as we know thread is not NULL, it is a JavaThread
>> and it has to have a
>>      >      >      non-null threadObj.
>>      >      >
>>      >      >      ---
>>      >      >
>>      >      > src/hotspot/share/services/management.cpp
>>      >      >
>>      >      >      1323         if (THREAD->is_Java_thread()) {
>>      >      >      1324           JavaThread* current_thread =
>> (JavaThread*)THREAD;
>>      >      >
>>      >      >      These calls can only be made on a JavaThread so
>> this be simplified to
>>      >      >      remove the is_Java_thread() call. Similarly in
>> other places.
>>      >      >
>>      >      >      ---
>>      >      >
>>      >      > src/hotspot/share/services/threadTable.cpp
>>      >      >
>>      >      >         55 class ThreadTableEntry : public
>> CHeapObj<mtInternal> {
>>      >      >         56   private:
>>      >      >         57     jlong _tid;
>>      >      >
>>      >      >      I believe hotspot style is to not indent the
>> access modifiers in C++
>>      >      >      class declarations, so the above would just be:
>>      >      >
>>      >      >         55 class ThreadTableEntry : public
>> CHeapObj<mtInternal> {
>>      >      >         56 private:
>>      >      >         57   jlong _tid;
>>      >      >
>>      >      >      etc.
>>      >      >
>>      >      >        60     ThreadTableEntry(jlong tid, JavaThread*
>> java_thread) :
>>      >      >        61 _tid(tid),_java_thread(java_thread) {}
>>      >      >
>>      >      >      line 61 should be indented as it continues line 60.
>>      >      >
>>      >      >         67 class ThreadTableConfig : public AllStatic {
>>      >      >         ...
>>      >      >         71     static uintx get_hash(Value const&
>> value, bool* is_dead) {
>>      >      >
>>      >      >      The is_dead parameter still bothers me here. I
>> can't make enough sense
>>      >      >      out of the template code in ConcurrentHashtable to
>> see why we have to
>>      >      >      have it, but I'm concerned that its very existence
>> means we perhaps
>>      >      >      should not be trying to extend CHT in this
>> context. ??
>>      >      >
>>      >      >        115   size_t start_size_log = size_log >
>> DefaultThreadTableSizeLog
>>      >      >        116   ? size_log : DefaultThreadTableSizeLog;
>>      >      >
>>      >      >      line 116 should be indented, though in this case I
>> think a better layout
>>      >      >      would be:
>>      >      >
>>      >      >        115   size_t start_size_log =
>>      >      >        116       size_log > DefaultThreadTableSizeLog ?
>> size_log :
>>      >      >      DefaultThreadTableSizeLog;
>>      >      >
>>      >      >        131 double ThreadTable::get_load_factor() {
>>      >      >        132   return (double)_items_count/_current_size;
>>      >      >        133 }
>>      >      >
>>      >      >      Not sure that is doing what you want/expect. It
>> will perform integer
>>      >      >      division and then cast that whole integer to a
>> double. If you want
>>      >      >      double arithmetic you need:
>>      >      >
>>      >      >      return ((double)_items_count)/_current_size;
>>      >      >
>>      >      >      180     jlong          _tid;
>>      >      >      181     uintx         _hash;
>>      >      >
>>      >      >      Nit: no need for all those spaces before the
>> variable name.
>>      >      >
>>      >      >        183     ThreadTableLookup(jlong tid)
>>      >      >        184     : _tid(tid), _hash(primitive_hash(tid)) {}
>>      >      >
>>      >      >      line 184 should be indented.
>>      >      >
>>      >      >      201     ThreadGet():_return(NULL) {}
>>      >      >
>>      >      >      Nit: need space after :
>>      >      >
>>      >      >        211    assert(_is_initialized, "Thread table is
>> not initialized");
>>      >      >        212   _has_work = false;
>>      >      >
>>      >      >      line 211 is indented one space too far.
>>      >      >
>>      >      >      229     ThreadTableEntry* entry = new
>> ThreadTableEntry(tid,java_thread);
>>      >      >
>>      >      >      Nit: need space after ,
>>      >      >
>>      >      >      252   return _local_table->remove(thread,lookup);
>>      >      >
>>      >      >      Nit: need space after ,
>>      >      >
>>      >      >      Thanks,
>>      >      >      David
>>      >      >      ------
>>      >      >
>>      >      >      > Bug:
>> https://bugs.openjdk.java.net/browse/JDK-8185005
>>      >      >      >
>>      >      >      > Thanks!
>>      >      >      > --Daniil
>>      >      >      >
>>      >      >      >
>>      >      >      > On 7/8/19, 3:24 PM, "Daniel D. Daugherty"
>> <[hidden email]> wrote:
>>      >      >      >
>>      >      >      >      On 6/29/19 12:06 PM, Daniil Titov wrote:
>>      >      >      >      > Hi Serguei and David,
>>      >      >      >      >
>>      >      >      >      > Serguei is right,
>> ThreadTable::find_thread(java_tid) cannot  return a JavaThread with
>> an unmatched java_tid.
>>      >      >      >      >
>>      >      >      >      > Please find a new version of the fix that
>> includes the changes Serguei suggested.
>>      >      >      >      >
>>      >      >      >      > Regarding the concern about the
>> maintaining the thread table when it may never even be queried, one of
>>      >      >      >      > the options could be to add ThreadTable
>> ::isEnabled flag, set it to "false" by default, and wrap the calls to
>> the thread table
>>      >      >      >      > in ThreadsSMRSupport add_thread() and
>> remove_thread() methods to check this flag.
>>      >      >      >      >
>>      >      >      >      > When
>> ThreadsList::find_JavaThread_from_java_tid() is called for the first
>> time it could check if ThreadTable ::isEnabled
>>      >      >      >      > Is on and if not then set it on and
>> populate the thread table with all existing threads from the thread
>> list.
>>      >      >      >
>>      >      >      >      I have the same concerns as David H. about
>> this new ThreadTable.
>>      >      >      > ThreadsList::find_JavaThread_from_java_tid() is
>> only called from code
>>      >      >      >      in
>> src/hotspot/share/services/management.cpp so I think that table
>>      >      >      >      needs to enabled and populated only if it
>> is going to be used.
>>      >      >      >
>>      >      >      >      I've taken a look at the webrev below and I
>> see that David has
>>      >      >      >      followed up with additional comments.
>> Before I do a crawl through
>>      >      >      >      code review for this, I would like to see
>> the ThreadTable stuff
>>      >      >      >      made optional and David's other comments
>> addressed.
>>      >      >      >
>>      >      >      >      Another possible optimization is for
>> callers of
>>      >      >      > find_JavaThread_from_java_tid() to save the
>> calling thread's
>>      >      >      >      tid value before they loop and if the
>> current tid == saved_tid
>>      >      >      >      then use the current JavaThread* instead of
>> calling
>>      >      >      > find_JavaThread_from_java_tid() to get the
>> JavaThread*.
>>      >      >      >
>>      >      >      >      Dan
>>      >      >      >
>>      >      >      >      >
>>      >      >      >      > Webrev:
>> https://cr.openjdk.java.net/~dtitov/8185005/webrev.02/
>>      >      >      >      > Bug:
>> https://bugs.openjdk.java.net/browse/JDK-8185005
>>      >      >      >      >
>>      >      >      >      > Thanks!
>>      >      >      >      > --Daniil
>>      >      >      >      >
>>      >      >      >      > From: <[hidden email]>
>>      >      >      >      > Organization: Oracle Corporation
>>      >      >      >      > Date: Friday, June 28, 2019 at 7:56 PM
>>      >      >      >      > To: Daniil Titov
>> <[hidden email]>, OpenJDK Serviceability
>> <[hidden email]>,
>> "[hidden email]"
>> <[hidden email]>, "[hidden email]"
>> <[hidden email]>
>>      >      >      >      > Subject: Re: RFR: 8185005: Improve
>> performance of ThreadMXBean.getThreadInfo(long ids[], int maxDepth)
>>      >      >      >      >
>>      >      >      >      > Hi Daniil,
>>      >      >      >      >
>>      >      >      >      > I have several quick comments.
>>      >      >      >      >
>>      >      >      >      > The indent in the hotspot c/c++ files has
>> to be 2, not 4.
>>      >      >      >      >
>>      >      >      >      >
>> https://cr.openjdk.java.net/~dtitov/8185005/webrev.01/src/hotspot/share/runtime/threadSMR.cpp.frames.html
>>      >      >      >      > 614 JavaThread*
>> ThreadsList::find_JavaThread_from_java_tid(jlong java_tid) const {
>>      >      >      >      >   615     JavaThread* java_thread =
>> ThreadTable::find_thread(java_tid);
>>      >      >      >      >   616     if (java_thread == NULL &&
>> java_tid == PMIMORDIAL_JAVA_TID) {
>>      >      >      >      >   617         //
>> ThreadsSMRSupport::add_thread() is not called for the primordial
>>      >      >      >      >   618         // thread. Thus, we find
>> this thread with a linear search and add it
>>      >      >      >      >   619         // to the thread table.
>>      >      >      >      >   620         for (uint i = 0; i <
>> length(); i++) {
>>      >      >      >      >   621 JavaThread* thread = thread_at(i);
>>      >      >      >      >   622             if
>> (is_valid_java_thread(java_tid,thread)) {
>>      >      >      >      >   623 ThreadTable::add_thread(java_tid,
>> thread);
>>      >      >      >      >   624 return thread;
>>      >      >      >      >   625             }
>>      >      >      >      >   626         }
>>      >      >      >      >   627     } else if (java_thread != NULL
>> && is_valid_java_thread(java_tid, java_thread)) {
>>      >      >      >      >   628         return java_thread;
>>      >      >      >      >   629     }
>>      >      >      >      >   630     return NULL;
>>      >      >      >      >   631 }
>>      >      >      >      >   632 bool
>> ThreadsList::is_valid_java_thread(jlong java_tid, JavaThread*
>> java_thread) {
>>      >      >      >      >   633     oop tobj =
>> java_thread->threadObj();
>>      >      >      >      >   634     // Ignore the thread if it
>> hasn't run yet, has exited
>>      >      >      >      >   635     // or is starting to exit.
>>      >      >      >      >   636     return (tobj != NULL &&
>> !java_thread->is_exiting() &&
>>      >      >      >      >   637 java_tid ==
>> java_lang_Thread::thread_id(tobj));
>>      >      >      >      >   638 }
>>      >      >      >      >
>>      >      >      >      >   615     JavaThread* java_thread =
>> ThreadTable::find_thread(java_tid);
>>      >      >      >      >
>>      >      >      >      >    I'd suggest to rename find_thread() to
>> find_thread_by_tid().
>>      >      >      >      >
>>      >      >      >      > A space is missed after the comma:
>>      >      >      >      >    622 if
>> (is_valid_java_thread(java_tid,thread)) {
>>      >      >      >      >
>>      >      >      >      > An empty line is needed before L632.
>>      >      >      >      >
>>      >      >      >      > The name 'is_valid_java_thread' looks
>> wrong (or confusing) to me.
>>      >      >      >      > Something like
>> 'is_alive_java_thread_with_tid()' would be better.
>>      >      >      >      > It'd better to list parameters in the
>> opposite order.
>>      >      >      >      >
>>      >      >      >      > The call to is_valid_java_thread() is
>> confusing:
>>      >      >      >      >     627 } else if (java_thread != NULL &&
>> is_valid_java_thread(java_tid, java_thread)) {
>>      >      >      >      >
>>      >      >      >      > Why would the call
>> ThreadTable::find_thread(java_tid) return a JavaThread with an
>> unmatched java_tid?
>>      >      >      >      >
>>      >      >      >      >
>>      >      >      >      > Thanks,
>>      >      >      >      > Serguei
>>      >      >      >      >
>>      >      >      >      > On 6/28/19, 9:40 PM, "David Holmes"
>> <[hidden email]> wrote:
>>      >      >      >      >
>>      >      >      >      >      Hi Daniil,
>>      >      >      >      >
>>      >      >      >      >      The definition and use of this
>> hashtable (yet another hashtable
>>      >      >      >      >      implementation!) will need careful
>> examination. We have to be concerned
>>      >      >      >      >      about the cost of maintaining it
>> when it may never even be queried. You
>>      >      >      >      >      would need to look at footprint cost
>> and performance impact.
>>      >      >      >      >
>>      >      >      >      >      Unfortunately I'm just about to
>> board a plane and will be out for the
>>      >      >      >      >      next few days. I will try to look at
>> this asap next week, but we will
>>      >      >      >      >      need a lot more data on it.
>>      >      >      >      >
>>      >      >      >      >      Thanks,
>>      >      >      >      >      David
>>      >      >      >      >
>>      >      >      >      > On 6/28/19 3:31 PM, Daniil Titov wrote:
>>      >      >      >      > Please review the change that improves
>> performance of ThreadMXBean MXBean methods returning the
>>      >      >      >      > information for specific threads. The
>> change introduces the thread table that uses ConcurrentHashTable
>>      >      >      >      > to store one-to-one the mapping between
>> the thread ids and JavaThread objects and replaces the linear
>>      >      >      >      > search over the thread list in
>> ThreadsList::find_JavaThread_from_java_tid(jlong tid) method with the
>> lookup
>>      >      >      >      > in the thread table.
>>      >      >      >      >
>>      >      >      >      > Testing: Mach5 tier1,tier2 and tier3
>> tests successfully passed.
>>      >      >      >      >
>>      >      >      >      > Webrev:
>> https://cr.openjdk.java.net/~dtitov/8185005/webrev.01/
>>      >      >      >      > Bug:
>> https://bugs.openjdk.java.net/browse/JDK-8185005
>>      >      >      >      >
>>      >      >      >      > Thanks!
>>      >      >      >      >
>>      >      >      >      > Best regards,
>>      >      >      >      > Daniil
>>      >      >      >      >
>>      >      >      >      >
>>      >      >      >      >
>>      >      >      >      >
>>      >      >      >      >
>>      >      >      >      >
>>      >      >      >      >
>>      >      >      >
>>      >      >      >
>>      >      >      >
>>      >      >      >
>>      >      >
>>      >      >
>>      >      >
>>      >
>>      >
>>      >
>>
>>