bottleneck by java.lang.Class.getAnnotations() - rebased patch

Peter Levart peter.levart at gmail.com
Mon Dec 10 14:59:26 UTC 2012 

On 12/10/2012 07:18 AM, David Holmes wrote:
> Hi Peter,
>
> Sorry for the delay on this.

Thank you for taking it into consideration. I can only imagine how busy 
you are with other things.

>
> Generally your VolatileData and my ReflectionHelper are doing a 
> similar job. But I agree with your reasoning that all of the cached 
> SoftReferences are likely to be cleared at once, and so a 
> SoftReference to a helper object with direct references, is more 
> effective than a direct reference to a helper object with 
> SoftReferences. My initial stance with this was very conservative as 
> the more change that is introduced the more uncertainty there is about 
> the impact.
>
> I say the above primarily because I think, if I am to proceed with 
> this, I will need to separate out the general reflection caching 
> changes from the annotation changes. 
> There are a number of reasons for this:
>
> First, I'm not at all familiar with the implementation of annotations 
> at the VM or Java level, and the recent changes in this area just 
> exacerbate my ignorance of the mechanics. So I don't feel qualified to 
> evaluate that aspect.
>
> Second, the bulk of the reflection caching code is simplified by the 
> fact that due to current constraints on class redefinition the caching 
> is effectively idempotent for fields/methods/constructors. But that is 
> not the case for annotations.

I think that on the Class-level these two aspects can be separated. But 
not on the Field/Method/Constructor level, because annotations are 
referenced by Field/Method/Constructor objects and there is no 
invalidation logic - like on the Class-level - that would just 
invalidate the sets of annotations on Field/Method/Constructor, leaving 
Field/Method/Constructor objects still valid. So these two aspects are 
connected - it may be - I haven't looked at history yet - that 
invalidation of Field/Method/Constructor was introduced just because of 
annotations.

Also, the following bug (currently not accessible): 
http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=5002251 has 
references to docs that suggest that current class redefinition can 
introduce new methods, so If this is correct, then redefinition is not 
idempotent for basic reflection data.

>
> Finally, the use of synchronization with the annotations method is 
> perplexing me. I sent Joe a private email on this but I may as well 
> raise it here - and I think you have alluded to this in your earlier 
> emails as well: initAnnotationsIfNecessary() is a synchronized 
> instance method but I can not find any other code in the VM that 
> synchronizes on the Class object's monitor. So if this synchronization 
> is trying to establish consistency in the face of class redefinition, 
> I do not see where class redefinition is participating in the 
> synchronization!

I think that the intent was merely synchronized access to / lazy 
initialization of cached 'annotations' and 'declaredAnnotations' Maps. 
Field[], Method[], Constructor[] arrays are published to other threads 
via volatile fields one field at a time, but 
initAnnotationsIfNecessary() was designed to publish two references 
('annotations' and 'declaredAnnotations') atomically at the same time, 
so the author of the code choose to use synchronized block. I also have 
a feeling that this might have simply been the author's preferred style 
of synchronization, since the same approach is used also in 
Field/Method/Constructor/Executable although there's just one field of 
annotations that is published at a time.

It is indicative that initAnnotationsIfNecessary() synchronized method 
also contains the call to clearCachesOnClassRedefinition(), so at first 
it seems that the synchronization is also meant to serialize access to 
invalidation logic which invalidates both Field/Method/Constructor and 
annotation fields, but that all falls-apart because 
clearCachesOnClassRedefinition() is also called from elsewhere, not 
guarded by the synchronization.

So all in all the two aspects - annotations and basic reflection stuff - 
are quite intermingled in current code, unfortunately very 
inconsistently. I'm afraid that doing one thing and not touching the 
other is possible, but that would not solve the problems that this 
thread was starting to address (bottleneck by 
java.lang.Class.getAnnotations()) and evident dead-lock bugs.

We can approach the problem so that we separate the aspects of caching 
Class-level annotations and Field/Method/Constructor arrays but using 
the same approach for both. And that would only make sense if there was 
a reason to separately cache Class-level annotations and 
Field/Method/Constructor arrays. I haven't yet been able to think of one 
such reason. So anyone with more insight (the author of annotations 
code?) is invited to participate in investigation.

My approach of including the Class-level annotations together with 
Field/Method/Constructor arrays was guided by:
- consistency - why should Class-level annotations be cached with hard 
references, while Field/Method/Constructor annotations are indirectly 
SoftReference(d)? Are they more important?
- simplicity
- space efficiency
- correctness - unsynchronized calls to clearCachesOnClassRedefinition() 
followed by unsynchronized lazy initialization have - as I have shown - 
theoretical races that could result in caching the old versions of data 
instead of new. The approach I have chosen with the logic in 
getVolatileData() is a kind of MVCC rather than synchronization.

But as said, the two aspects of caching can be separated.

We can also leave the Class-level annotation aspect untouched by 
re-introducing the 'annotations' and 'declaredAnnotations' fields and 
also the 'lastRedefinedCount' field on the Class-level and 
re-introducing the synchronized initAnnotationsIfNecessary() method and 
a clearCachesOnClassRedefinition() which would just invalidate the two 
annotations fields.

To recap. How to continue?
a) as proposed;
b) separate caching of annotations and Field/Method/Constructor arrays 
but with the same unblocking MVCC-like approach for both - with possible 
variation in that annotations be hardly referenced and 
Field/Method/Constructor arrays be softly;
c) undo the annotations caching changes and only do MVCC for 
Field/Method/Constructor arrays.

I can do b) but prepare two independent patches - one for VolatileData 
(rename it to MemberData?) and one for AnnotationData. So by applying 
only the first, we achieve c) and can later apply the second to upgrade 
to b). But it is unfortunately a) that is most efficient space-saving wise.

What do you say about the trivial changes in 
Field/Method/Constructor/Executable to accommodate caching on the 'root' 
instance instead of on the copies?

Regards, Peter

>
> So what I would like to do is take your basic VolatileData part of the 
> patch and run with it for JEP-149 purposes, while separating the 
> annotations issue so they can be dealt with by the experts in that 
> particular area.
>
> I'm sorry it has taken so long to arrive at a fairly negative 
> position, but I need someone else to take up the annotations gauntlet 
> and run with it.
>
> Thanks,
> David
>
> On 3/12/2012 5:41 PM, Peter Levart wrote:
>> Hi David, Alan, Alexander and others,
>>
>> In the meanwhile I have added another annotations space optimization to
>> the patch. If a Class doesn't inherit any annotations from a superclass,
>> which I think is a common case, it assigns the same Map instance to
>> "annotations" as well as "declaredAnnotations" fields. Previously - and
>> in the original code - this only happened for java.lang.Object and
>> interfaces.
>>
>> Here's the updated webrev:
>>
>> http://dl.dropbox.com/u/101777488/jdk8-tl/JEP-149/webrev.02/index.html
>>
>> I have also rewritten the performance micro-benchmarks. With the
>> addition of repeating annotations, one performance aspect surfaces: when
>> asking for a particular annotation type on a Class and that annotation
>> is not present, the new repeating annotations support method
>> AnnotationSupport.getOneAnnotation asks for @ContainedBy meta-annotation
>> on the annotation type. This can result in an even more apparent
>> synchronization hot-spot with original code that uses synchronized
>> initAnnotationsIfNecessary(). This aspect is tested with the 3rd test.
>> Other 2 tests test the same thing as before but are more stable now,
>> since now they measure retrieval of 5 different annotation types from
>> each AnnotatedElement, previously they only measured retrieval of 1
>> which was very sensitive to HashMap irregularities (it could happen that
>> a particular key mapped to a bucket that was overloaded in one test-run
>> and not in another)...
>>
>> Here're the new tests:
>>
>> https://raw.github.com/plevart/jdk8-tl/JEP-149/test/src/test/ReflectionTest.java 
>>
>>
>> And the corresponding results when run on an i7 CPU on Linux:
>>
>> https://raw.github.com/plevart/jdk8-tl/JEP-149/test/benchmark_results_i7-2600K.txt 
>>
>>
>>
>> Regards, Peter
>>
>>
>>
>> On 12/03/2012 02:15 AM, David Holmes wrote:
>>> On 1/12/2012 4:54 AM, Alan Bateman wrote:
>>>> On 30/11/2012 18:36, Peter Levart wrote:
>>>>> :
>>>>>
>>>>> So, what do you think? What kind of tests should I prepare in 
>>>>> addidion
>>>>> to those 3 so that the patch might get a consideration?
>>>> I think this is good work and thanks for re-basing your patch. I know
>>>> David plans to do a detail review. I think it will require extensive
>>>> performance testing too, perhaps with some large applications.
>>>
>>> Indeed I do plan a detailed review and have initiated some initial
>>> performance tests.
>>>
>>> I am also swamped but will try to get to the review this week - and
>>> will also need to check the referenced annotations updates.
>>>
>>> David
>>>
>>>> -Alan
>>>>
>>

More information about the core-libs-dev mailing list