JDK-8171119: Low-Overhead Heap Profiling

White, Derek Derek.White at cavium.com
Tue Apr 3 22:54:07 UTC 2018

Thanks JC,

New patch applies cleanly. Compiles and runs (simple test programs) on aarch64.

  *   Derek

From: JC Beyler [mailto:jcbeyler at google.com]
Sent: Monday, April 02, 2018 1:17 PM
To: White, Derek <Derek.White at cavium.com>
Cc: Erik Österlund <erik.osterlund at oracle.com>; serviceability-dev at openjdk.java.net; hotspot-compiler-dev <hotspot-compiler-dev at openjdk.java.net>
Subject: Re: JDK-8171119: Low-Overhead Heap Profiling

Hi Derek,

I know there were a few things that went in that provoked a merge conflict. I worked on it and got it up to date. Sadly my lack of knowledge makes it a full rebase instead of keeping all the history. However, with a newly cloned jdk/hs you should now be able to use:

The change you are referring to was done with the others so perhaps you were unlucky and I forgot it in a webrev and fixed it in another? I don't know but it's been there and I checked, it is here:

I double checked that tlab_end_offset no longer appears in any architecture (as far as I can tell :)).

Thanks for testing and let me know if you run into any other issues!

On Fri, Mar 30, 2018 at 4:24 PM White, Derek <Derek.White at cavium.com<mailto:Derek.White at cavium.com>> wrote:
Hi Jc,

I’ve been having trouble getting your patch to apply correctly. I may have based it on the wrong version.

In any case, I think there’s a missing update to macroAssembler_aarch64.cpp, in MacroAssembler::tlab_allocate(), where “JavaThread::tlab_end_offset()” should become “JavaThread::tlab_current_end_offset()”.

This should correspond to the other port’s changes in templateTable_<cpu>.cpp files.

- Derek

From: hotspot-compiler-dev [mailto:hotspot-compiler-dev-bounces at openjdk.java.net<mailto:hotspot-compiler-dev-bounces at openjdk.java.net>] On Behalf Of JC Beyler
Sent: Wednesday, March 28, 2018 11:43 AM
To: Erik Österlund <erik.osterlund at oracle.com<mailto:erik.osterlund at oracle.com>>
Cc: serviceability-dev at openjdk.java.net<mailto:serviceability-dev at openjdk.java.net>; hotspot-compiler-dev <hotspot-compiler-dev at openjdk.java.net<mailto:hotspot-compiler-dev at openjdk.java.net>>
Subject: Re: JDK-8171119: Low-Overhead Heap Profiling

Hi all,

I've been working on deflaking the tests mostly and the wording in the JVMTI spec.

Here is the two incremental webrevs:

Here is the total webrev:

Here are the notes of this change:
  - Currently the tests pass 100 times in a row, I am working on checking if they pass 1000 times in a row.
  - The default sampling rate is set to 512k, this is what we use internally and having a default means that to enable the sampling with the default, the user only has to do a enable event/disable event via JVMTI (instead of enable + set sample rate).
  - I deprecated the code that was handling the fast path tlab refill if it happened since this is now deprecated
      - Though I saw that Graal is still using it so I have to see what needs to be done there exactly

Finally, using the Dacapo benchmark suite, I noted a 1% overhead for when the event system is turned on and the callback to the native agent is just empty. I got a 3% overhead with a 512k sampling rate with the code I put in the native side of my tests.

Thanks and comments are appreciated,

On Mon, Mar 19, 2018 at 2:06 PM JC Beyler <jcbeyler at google.com<mailto:jcbeyler at google.com>> wrote:
Hi all,

The incremental webrev update is here:

The full webrev is here:

Major change here is:
  - I've removed the heapMonitoring.cpp code in favor of just having the sampling events as per Serguei's request; I still have to do some overhead measurements but the tests prove the concept can work
       - Most of the tlab code is unchanged, the only major part is that now things get sent off to event collectors when used and enabled.
  - Added the interpreter collectors to handle interpreter execution
  - Updated the name from SetTlabHeapSampling to SetHeapSampling to be more generic
  - Added a mutex for the thread sampling so that we can initialize an internal static array safely
  - Ported the tests from the old system to this new one

I've also updated the JEP and CSR to reflect these changes:

In order to make this have some forward progress, I've removed the heap sampling code entirely and now rely entirely on the event sampling system. The tests reflect this by using a simplified implementation of what an agent could do:
(Search for anything mentioning event_storage).

I have not taken the time to port the whole code we had originally in heapMonitoring to this. I hesitate only because that code was in C++, I'd have to port it to C and this is for tests so perhaps what I have now is good enough?

As far as testing goes, I've ported all the relevant tests and then added a few:
   - Turning the system on/off
   - Testing using various GCs
   - Testing using the interpreter
   - Testing the sampling rate
   - Testing with objects and arrays
   - Testing with various threads

Finally, as overhead goes, I have the numbers of the system off vs a clean build and I have 0% overhead, which is what we'd want. This was using the Dacapo benchmarks. I am now preparing to run a version with the events on using dacapo and will report back here.

Any comments are welcome :)

On Thu, Mar 8, 2018 at 4:00 PM JC Beyler <jcbeyler at google.com<mailto:jcbeyler at google.com>> wrote:
Hi all,

I apologize for the delay but I wanted to add an event system and that took a bit longer than expected and I also reworked the code to take into account the deprecation of FastTLABRefill.

This update has four parts:

A) I moved the implementation from Thread to ThreadHeapSampler inside of Thread. Would you prefer it as a pointer inside of Thread or like this works for you? Second question would be would you rather have an association outside of Thread altogether that tries to remember when threads are live and then we would have something like:

I worry about the overhead of this but perhaps it is not too too bad?

B) I also have been working on the Allocation event system that sends out a notification at each sampled event. This will be practical when wanting to do something at the allocation point. I'm also looking at if the whole heapMonitoring code could not reside in the agent code and not in the JDK. I'm not convinced but I'm talking to Serguei about it to see/assess :)
   - Also added two tests for the new event subsystem

C) Removed the slow_path fields inside the TLAB code since now FastTLABRefill is deprecated

D) Updated the JVMTI documentation and specification for the methods.

So the incremental webrev is here:

and the full webrev is here:

I believe I have updated the various JIRA issues that track this :)

Thanks for your input,

On Wed, Feb 14, 2018 at 10:34 PM, JC Beyler <jcbeyler at google.com<mailto:jcbeyler at google.com>> wrote:
Hi Erik,

I inlined my answers, which the last one seems to answer Robbin's concerns about the same thing (adding things to Thread).

On Wed, Feb 14, 2018 at 2:51 AM, Erik Österlund <erik.osterlund at oracle.com<mailto:erik.osterlund at oracle.com>> wrote:
Hi JC,

Comments are inlined below.

On 2018-02-13 06:18, JC Beyler wrote:
Hi Erik,

Thanks for your answers, I've now inlined my own answers/comments.

I've done a new webrev here:

The incremental is here:

Note to all:
  - I've been integrating changes from Erin/Serguei/David comments so this webrev incremental is a bit an answer to all comments in one. I apologize for that :)

On Mon, Feb 12, 2018 at 6:05 AM, Erik Österlund <erik.osterlund at oracle.com<mailto:erik.osterlund at oracle.com>> wrote:
Hi JC,

Sorry for the delayed reply.

Inlined answers:

On 2018-02-06 00:04, JC Beyler wrote:
Hi Erik,

(Renaming this to be folded into the newly renamed thread :))

First off, thanks a lot for reviewing the webrev! I appreciate it!

I updated the webrev to:

And the incremental one is here:

It contains:
- The change for since from 9 to 11 for the jvmti.xml
- The use of the OrderAccess for initialized
- Clearing the oop

I also have inlined my answers to your comments. The biggest question
will come from the multiple *_end variables. A bit of the logic there
is due to handling the slow path refill vs fast path refill and
checking that the rug was not pulled underneath the slowpath. I
believe that a previous comment was that TlabFastRefill was going to
be deprecated.

If this is true, we could revert this code a bit and just do a : if
TlabFastRefill is enabled, disable this. And then deprecate that when
TlabFastRefill is deprecated.

This might simplify this webrev and I can work on a follow-up that
either: removes TlabFastRefill if Robbin does not have the time to do
it or add the support to the assembly side to handle this correctly.
What do you think?

I support removing TlabFastRefill, but I think it is good to not depend on that happening first.

I'm slowly pushing on the FastTLABRefill (https://bugs.openjdk.java.net/browse/JDK-8194084), I agree on keeping both separate for now though so that we can think of both differently

Now, below, inlined are my answers:

On Fri, Feb 2, 2018 at 8:44 AM, Erik Österlund
<erik.osterlund at oracle.com<mailto:erik.osterlund at oracle.com>> wrote:
Hi JC,

Hope I am reviewing the right version of your work. Here goes...


  159     AllocTracer::send_allocation_outside_tlab(klass, result, size *
HeapWordSize, THREAD);
  161     THREAD->tlab().handle_sample(THREAD, result, size);
  162     return result;
  163   }

Should not call tlab()->X without checking if (UseTLAB) IMO.

More about this later.


So first of all, there seems to quite a few ends. There is an "end", a "hard
end", a "slow path end", and an "actual end". Moreover, it seems like the
"hard end" is actually further away than the "actual end". So the "hard end"
seems like more of a "really definitely actual end" or something. I don't
know about you, but I think it looks kind of messy. In particular, I don't
feel like the name "actual end" reflects what it represents, especially when
there is another end that is behind the "actual end".

  413 HeapWord* ThreadLocalAllocBuffer::hard_end() {
  414   // Did a fast TLAB refill occur?
  415   if (_slow_path_end != _end) {
  416     // Fix up the actual end to be now the end of this TLAB.
  417     _slow_path_end = _end;
  418     _actual_end = _end;
  419   }
  421   return _actual_end + alignment_reserve();
  422 }

I really do not like making getters unexpectedly have these kind of side
effects. It is not expected that when you ask for the "hard end", you
implicitly update the "slow path end" and "actual end" to new values.
As I said, a lot of this is due to the FastTlabRefill. If I make this
not supporting FastTlabRefill, this goes away. The reason the system
needs to update itself at the get is that you only know at that get if
things have shifted underneath the tlab slow path. I am not sure of
really better names (naming is hard!), perhaps we could do these

- current_tlab_end       // Either the allocated tlab end or a sampling point
- last_allocation_address  // The end of the tlab allocation
- last_slowpath_allocated_end  // In case a fast refill occurred the
end might have changed, this is to remember slow vs fast past refills

the hard_end method can be renamed to something like:
tlab_end_pointer()        // The end of the lab including a bit of
alignment reserved bytes

Those names sound better to me. Could you please provide a mapping from the old names to the new names so I understand which one is which please?

This is my current guess of what you are proposing:

end -> current_tlab_end
actual_end -> last_allocation_address
slow_path_end -> last_slowpath_allocated_end
hard_end -> tlab_end_pointer

Yes that is correct, that was what I was proposing.

I would prefer this naming:

end -> slow_path_end // the end for taking a slow path; either due to sampling or refilling
actual_end -> allocation_end // the end for allocations
slow_path_end -> last_slow_path_end // last address for slow_path_end (as opposed to allocation_end)
hard_end -> reserved_end // the end of the reserved space of the TLAB

About setting things in the getter... that still seems like a very unpleasant thing to me. It would be better to inspect the call hierarchy and explicitly update the ends where they need updating, and assert in the getter that they are in sync, rather than implicitly setting various ends as a surprising side effect in a getter. It looks like the call hierarchy is very small. With my new naming convention, reserved_end() would presumably return _allocation_end + alignment_reserve(), and have an assert checking that _allocation_end == _last_slow_path_allocation_end, complaining that this invariant must hold, and that a caller to this function, such as make_parsable(), must first explicitly synchronize the ends as required, to honor that invariant.

I've renamed the variables to how you preferred it except for the _end one. I did:

The reason is that the architecture dependent code use the thread.hpp API and it already has tlab included into the name so it becomes tlab_current_end (which is better that tlab_current_tlab_end in my opinion).

I also moved the update into a separate method with a TODO that says to remove it when FastTLABRefill is deprecated

This looks a lot better now. Thanks.

Note that the following comment now needs updating accordingly in threadLocalAllocBuffer.hpp:

  41 //            Heap sampling is performed via the end/actual_end fields.

  42 //            actual_end contains the real end of the tlab allocation,

  43 //            whereas end can be set to an arbitrary spot in the tlab to

  44 //            trip the return and sample the allocation.

  45 //            slow_path_end is used to track if a fast tlab refill occured

  46 //            between slowpath calls.
There might be other comments too, I have not looked in detail.

This was the only spot that still had an actual_end, I fixed it now. I'll do a sweep to double check other comments.

Not sure it's better but before updating the webrev, I wanted to try
to get input/consensus :)

(Note hard_end was always further off than end).

10357       <capabilityfield id="can_sample_heap" since="9">
10358         <description>
10359           Can sample the heap.
10360           If this capability is enabled then the heap sampling methods
can be called.
10361         </description>
10362       </capabilityfield>

Looks like this capability should not be "since 9" if it gets integrated
Updated now to 11, crossing my fingers :)

  448       if (is_alive->do_object_b(value)) {
  449         // Update the oop to point to the new object if it is still
  450         f->do_oop(&(trace.obj));
  452         // Copy the old trace, if it is still live.
  453         _allocated_traces->at_put(curr_pos++, trace);
  455         // Store the live trace in a cache, to be served up on /heapz.
  456         _traces_on_last_full_gc->append(trace);
  458         count++;
  459       } else {
  460         // If the old trace is no longer live, add it to the list of
  461         // recently collected garbage.
  462         store_garbage_trace(trace);
  463       }

In the case where the oop was not live, I would like it to be explicitly
Done I think how you wanted it. Let me know because I'm not familiar
with the RootAccess API. I'm unclear if I'm doing this right or not so
reviews of these parts are highly appreciated. Robbin had talked of
perhaps later pushing this all into a OopStorage, should I do this now
do you think? Or can that wait a second webrev later down the road?

I think using handles can and should be done later. You can use the Access API now.
I noticed that you are missing an #include "oops/access.inline.hpp" in your heapMonitoring.cpp file.

The missing header is there for me so I don't know, I made sure it is present in the latest webrev. Sorry about that.

+ Did I clear it the way you wanted me to or were you thinking of
something else?

That is precisely how I wanted it to be cleared. Thanks.
+ Final question here, seems like if I were to want to not do the
f->do_oop directly on the trace.obj, I'd need to do something like:


to update the oop internally. Is that right/is that one of the
advantages of going to the Oopstorage sooner than later?

I think you really want to do the do_oop on the root directly. Is there a particular reason why you would not want to do that?
Otherwise, yes - the benefit with using the handle approach is that you do not need to call do_oop explicitly in your code.

There is no reason except that now we have a load_oop and a get_oop_addr, I was not sure what you would think of that.

That's fine.

Also I see a lot of concurrent-looking use of the following field:
  267   volatile bool _initialized;

Please note that the "volatile" qualifier does not help with reordering
here. Reordering between volatile and non-volatile fields is completely free
for both compiler and hardware, except for windows with MSVC, where volatile
semantics is defined to use acquire/release semantics, and the hardware is
TSO. But for the general case, I would expect this field to be stored with
OrderAccess::release_store and loaded with OrderAccess::load_acquire.
Otherwise it is not thread safe.
Because everything is behind a mutex, I wasn't really worried about
this. I have a test that has multiple threads trying to hit this
corner case and it passes.

However, to be paranoid, I updated it to using the OrderAccess API
now, thanks! Let me know what you think there too!

If it is indeed always supposed to be read and written under a mutex, then I would strongly prefer to have it accessed as a normal non-volatile member, and have an assertion that given lock is held or we are in a safepoint, as we do in many other places. Something like this:

assert(HeapMonitorStorage_lock->owned_by_self() || (SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread()), "this should not be accessed concurrently");

It would be confusing to people reading the code if there are uses of OrderAccess that are actually always protected under a mutex.

Thank you for the exact example to be put in the code! I put it around each access/assignment of the _initialized method and found one case where yes you can touch it and not have the lock. It actually is "ok" because you don't act on the storage until later and only when you really want to modify the storage (see the object_alloc_do_sample method which calls the add_trace method).

But, because of this, I'm going to put the OrderAccess here, I'll do some performance numbers later and if there are issues, I might add a "unsafe" read and a "safe" one to make it explicit to the reader. But I don't think it will come to that.

Okay. This double return in heapMonitoring.cpp looks wrong:

 283   bool initialized() {
 284     return OrderAccess::load_acquire(&_initialized) != 0;
 285     return _initialized;
 286   }

Since you said object_alloc_do_sample() is the only place where you do not hold the mutex while reading initialized(), I had a closer look at that. It looks like in its current shape, the lack of a mutex may lead to a memory leak. In particular, it first checks if (initialized()). Let's assume this is now true. It then allocates a bunch of stuff, and checks if the number of frames were over 0. If they were, it calls StackTraceStorage::storage()->add_trace() seemingly hoping that after grabbing the lock in there, initialized() will still return true. But it could now return false and skip doing anything, in which case the allocated stuff will never be freed.

I fixed this now by making add_trace return a boolean and checking for that. It will be in the next webrev. Thanks, the truth is that in our implementation the system is always on or off, so this never really occurs :). In this version though, that is not true and it's important to handle so thanks again!

So the analysis seems to be that _initialized is only used outside of the mutex in once instance, where it is used to perform double-checked locking, that actually causes a memory leak.

I am not proposing how to fix that, just raising the issue. If you still want to perform this double-checked locking somehow, then the use of acquire/release still seems odd. Because the memory ordering restrictions of it never comes into play in this particular case. If it ever did, then the use of destroy_stuff(); release_store(_initialized, 0) would be broken anyway as that would imply that whatever concurrent reader there ever was would after reading _initialized with load_acquire() could *never* read the data that is concurrently destroyed anyway. I would be biased to think that RawAccess<MO_RELAXED>::load/store looks like a more appropriate solution, given that the memory leak issue is resolved. I do not know how painful it would be to not perform this double-checked locking.

So I agree with this entirely. I looked also a bit more and the difference and code really stems from our internal version. In this version however, there are actually a lot of things going on that I did not go entirely through in my head but this comment made me ponder a bit more on it.

Since every object_alloc_do_sample is protected by a check to HeapMonitoring::enabled(), there is only a small chance that the call is happening when things have been disabled. So there is no real need to do a first check on the initialized, it is a rare occurence that a call happens to object_alloc_do_sample and the initialized of the storage returns false.

(By the way, even if you did call object_alloc_do_sample without looking at HeapMonitoring::enabled(), that would be ok too. You would gather the stacktrace and get nowhere at the add_trace call, which would return false; so though not optimal performance wise, nothing would break).

Furthermore, the add_trace is really the moment of no return and we have the mutex lock and then the initialized check. So, in the end, I did two things: I removed that first check and then I removed the OrderAccess for the storage initialized. I think now I have a better grasp and understanding why it was done in our code and why it is not needed here. Thanks for pointing it out :). This now still passes my JTREG tests, especially the threaded one.

As a kind of meta comment, I wonder if it would make sense to add sampling
for non-TLAB allocations. Seems like if someone is rapidly allocating a
whole bunch of 1 MB objects that never fit in a TLAB, I might still be
interested in seeing that in my traces, and not get surprised that the
allocation rate is very high yet not showing up in any profiles.
That is handled by the handle_sample where you wanted me to put a
UseTlab because you hit that case if the allocation is too big.

I see. It was not obvious to me that non-TLAB sampling is done in the TLAB class. That seems like an abstraction crime.
What I wanted in my previous comment was that we do not call into the TLAB when we are not using TLABs. If there is sampling logic in the TLAB that is used for something else than TLABs, then it seems like that logic simply does not belong inside of the TLAB. It should be moved out of the TLAB, and instead have the TLAB call this common abstraction that makes sense.

So in the incremental version:
http://cr.openjdk.java.net/~jcbeyler/8171119/webrev.07_08/<http://cr.openjdk.java.net/%7Ejcbeyler/8171119/webrev.07_08/>, this is still a "crime". The reason is that the system has to have the bytes_until_sample on a per-thread level and it made "sense" to have it with the TLAB implementation. Also, I was not sure how people felt about adding something to the thread instance instead.

Do you think it fits better at the Thread level? I can see how difficult it is to make it happen there and add some logic there. Let me know what you think.

We have an unfortunate situation where everyone that has some fields that are thread local tend to dump them right into Thread, making the size and complexity of Thread grow as it becomes tightly coupled with various unrelated subsystems. It would be desirable to have a separate class for this instead that encapsulates the sampling logic. That class could possibly reside in Thread though as a value object of Thread.

I imagined that would be the case but was not sure. I will look at the example that Robbin is talking about (ThreadSMR) and will see how to refactor my code to use that.

Thanks again for your help,

Hope I have answered your questions and that my feedback makes sense to you.

You have and thank you for them, I think we are getting to a cleaner implementation and things are getting better and more readable :)

Yes it is getting better.


Thanks for your help!


I double checked by changing the test

to use a smaller Tlab (2048) and made the object bigger and it goes
through that and passes.

Thanks again for your review and I look forward to your pointers for
the questions I now have raised!


On 2018-01-26 06:45, JC Beyler wrote:
Thanks Robbin for the reviews :)

The new full webrev is here:
The incremental webrev is here:

I inlined my answers:

On Thu, Jan 25, 2018 at 1:15 AM, Robbin Ehn <robbin.ehn at oracle.com<mailto:robbin.ehn at oracle.com>> wrote:
Hi JC, great to see another revision!


StackTraceData should not contain the oop for 'safety' reasons.
When StackTraceData is moved from _allocated_traces:
L452 store_garbage_trace(trace);
it contains a dead oop.
_allocated_traces could instead be a tupel of oop and StackTraceData thus
dead oops are not kept.
Done I used inheritance to make the copier work regardless but the
idea is the same.
You should use the new Access API for loading the oop, something like
I don't think you need to use Access API for clearing the oop, but it
look nicer. And you shouldn't probably be using:
I am unfamiliar with this but I think I did do it like you wanted me
to (all tests pass so that's a start). I'm not sure how to clear the
oop exactly, is there somewhere that does that, which I can use to do
the same?

I removed the is_in_reserved, this came from our internal version, I
don't know why it was there but my tests work without so I removed it
The lock:
L424   MutexLocker mu(HeapMonitorStorage_lock);
Is not needed as far as I can see.
weak_oops_do is called in a safepoint, no TLAB allocation can happen and
JVMTI thread can't access these data-structures. Is there something more
this lock that I'm missing?
Since a thread can call the JVMTI getLiveTraces (or any of the other
ones), it can get to the point of trying to copying the
_allocated_traces. I imagine it is possible that this is happening
during a GC or that it can be started and a GC happens afterwards.
Therefore, it seems to me that you want this protected, no?
You have 6 files without any changes in them (any more):
I have not looked closely, but is it possible to hide heap sampling in
AllocTracer ? (with some minor changes to the AllocTracer API)
I am imagining that you are saying to move the code that does the
sampling code (change the tlab end, do the call to HeapMonitoring,
etc.) into the AllocTracer code itself? I think that is right and I'll
look if that is possible and prepare a webrev to show what would be
needed to make that happen.
Minor nit, when declaring pointer there is a little mix of having the
pointer adjacent by type name and data name. (Most hotspot code is by
heapMonitoring.cpp:711     jvmtiStackTrace *trace = ....
heapMonitoring.cpp:733         Method* m = vfst.method();
(not just this file)
I would make g_tmp volatile, otherwise the assignment in loop may
theoretical be skipped.
Also done!

Thanks again!

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openjdk.java.net/pipermail/hotspot-compiler-dev/attachments/20180403/a0f39ae4/attachment-0001.html>

More information about the hotspot-compiler-dev mailing list