RFR: JDK-8055845 - Add trace event for promoted objects

Staffan Friberg staffan.friberg at oracle.com
Thu Nov 6 10:12:22 UTC 2014


Hi,

After further off list discussion it was decided to keep the gc_tracer 
in par_promote as is.

I have uploaded a new webrev, 
http://cr.openjdk.java.net/~sfriberg/8055845/webrev.05

The main change here is a rewrite of the G1 code which is cleaner and 
also reuses the read age. By sending the markOop down through the call 
we can now trust the read age and do not need to reread it when 
incrementing which improves the YC performance slightly as it avoids the 
rather complex bit extraction.

Cheers,
Staffan

On 09/15/2014 02:06 PM, Bengt Rutisson wrote:
>
> Hi Staffan,
>
> psPromotionManager.inline.hpp
>
> I think the PSPromotionManager::copy_to_survivor_space() might send 
> multiple events. If the allocation to the young gen fails we will fall 
> through to do an old gen allocation. But we send the events before we 
> realize that the allocation has failed, i.e. new_obj is NULL.
>
> I talked to Erik a bit about how to handle the gc_tracer in 
> par_promote. He'll get back to you with some thoughts on that.
>
> Thanks,
> Bengt
>
>
> On 2014-09-06 00:20, Staffan Friberg wrote:
>> Hi,
>>
>> I have uploaded a new webrev here, 
>> cr.openjdk.java.net/~sfriberg/8055845/webrev.03
>>
>> It contains several changes
>>
>>     - Split event into two events (PromoteObjectInNewPLAB, 
>> PromoteObjectOutsidePLAB)
>>     - Moved events to "vm/gc/detailed/PromoteObject..."
>>     - Supporting ParNew+CMS and ParNew+SerialOld tenuring
>>          - Not sure if the way I do it with passing the ParNewTracer 
>> is the best solution, please let me know if you have an idea how to 
>> improve it
>>     - Simplified the G1 code to avoid sending age and having a single 
>> call site
>>     - Fixed so that the generated event has size information in bytes 
>> rather than words
>>
>> Thanks for offline comments and suggestions from Dmitry and Thomas.
>>
>> Cheers,
>> Staffan
>>
>> On 08/29/2014 03:32 PM, Staffan Friberg wrote:
>>> Hi Erik,
>>>
>>> On 08/28/2014 11:34 PM, Erik Helin wrote:
>>>> (it seems like we lost hotspot-gc-dev at openjdk.java.net somewhere in 
>>>> this thread, I'm adding it back)
>>>>
>>>> On 2014-08-28 23:15, Staffan Friberg wrote:
>>>>> Hi Erik,
>>>>>
>>>>> Thanks for the comments.
>>>>>> - Aren't the events for promotion to the tenured generation 
>>>>>> (SerialOld)
>>>>>>   and the CMS generation missing?
>>>>> The reason for leaving out these two, Serial completely and CMS
>>>>> promotion, was due to that neither as far as I understand make use of
>>>>> PLABs.
>>>>
>>>> I might be wrong here, but looking at the function 
>>>> TenuredGeneration::par_promote (in tenuredGeneration.cpp) it looks 
>>>> to me like SerialOld is using PLABs when ParNew is promoting 
>>>> objects from young to old.
>>>>
>>>> As for CMS, looking at ConcurrentMarkSweepGeneration::par_promote 
>>>> (in concurrentMarkSweepGeneration.cpp) it seems like each 
>>>> CMSParGCThreadState has a CFLS_LAB (CompactibleFreeListSpace Local 
>>>> Allocation Buffer) that is a thread-local allocation buffer. See 
>>>> compactibleFreeListSpace.{hpp,cpp} for more details.
>>>>
>>>> Given this, I think we should add events for Serial and CMS as well.
>>>>
>>>
>>> Ok I see what you mean with CMS, basically the equivalent to getting 
>>> a PLAB would be when we refill the CFLS_LAB in the alloc function. 
>>> It still does the allocation to a small memory (~ size of object) 
>>> area from the freelist, but at least we did extra work to refill the 
>>> local CFLS_LAB. Need to do some analysis to see how often we refill 
>>> so the overhead doesn't get too high.
>>> The only issue I run into is how I can in a nice way get access to 
>>> the ParNewTracer from ParNewGeneration to call the report function. 
>>> Let's sync up next week and see how it can be solved.
>>>
>>> The tenured GC requires something similar to be supported, however 
>>> since ParNewGC is deprecated for usage without CMS in JDK 8 we might 
>>> want to skip that combination.
>>>
>>>
>>>>
>>>> On 2014-08-28 23:15, Staffan Friberg wrote:
>>>>>> - Would it make sense to differentiate, either by separate events 
>>>>>> or by
>>>>>>   a field in a event, between promotions to to-space and to the old
>>>>>>   generation?
>>>>>> - The are two events for TLAB allocations,
>>>>>>   java/object_alloc_in_new_TLAB and java/object_alloc_outside_TLAB.
>>>>>>   What do you think about using two events for PLAB allocations 
>>>>>> as well:
>>>>>>   - java/object_alloc_in_new_PLAB
>>>>>>   - java/object_alloc_outside_PLAB
>>>>> I think this is a matter of taste and probably how similar we want 
>>>>> the
>>>>> event to be to the existing allocation event. I personally prefer a
>>>>> single event but if the GC team and serviceability team feel it 
>>>>> would be
>>>>> better to have two I can certainly rewrite. The reason for me 
>>>>> preferring
>>>>> a single event is just ease of analysis, you can easily filter a 
>>>>> list of
>>>>> events on a field, it is harder to merge two different events with
>>>>> different fields and work with them in an straight forward fashion.
>>>>>
>>>>> Any general preference for having a single or multiple events?
>>>>
>>>> I would prefer to have two events in this case and try to follow 
>>>> the existing allocation events as much as possible (both in naming 
>>>> and in style). Keeping the tenured field (I missed it the first 
>>>> time I read the patch) is good.
>>>>
>>> Yes, tenured would be independent of having two events, only PLAB 
>>> size and directAllocation would be affected when having two event types.
>>>
>>> *Erik Gahlin*, any preference from your end?
>>>
>>>
>>>
>>>> On 2014-08-28 23:15, Staffan Friberg wrote:
>>>>>> - In PSPromotionManager, instead of utilizing the C++ friendship 
>>>>>> with
>>>>>>   PSScavenge, should we add a getter function for the gc_tracer?
>>>>> Created a getter function.
>>>>
>>>> Thanks! If you make report_promotion_sample const, then the getter 
>>>> can return a const ParallelScavengeTracer*, right?
>>>>
>>> Done
>>>
>>> //Staffan
>>
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openjdk.java.net/pipermail/hotspot-gc-dev/attachments/20141106/eadbcac0/attachment-0001.html>


More information about the hotspot-gc-dev mailing list