bottleneck by java.lang.Class.getAnnotations() - a better patch

Peter Levart peter.levart at
Wed Nov 7 20:49:31 UTC 2012

Hi all,

I have redone the static memory footprint comparison calculations 
(taking correct object headers and padding into account) and I hope this 
time I've done it right. Here's what I got:

64bit addressing (16 byte object header):

patched Class uses 76 bytes less than original Class when empty
patched Class uses 360 bytes less than original Class when fully loaded
patched Class uses 32 bytes more than original Class when just one of 
fields is loaded

32bit addressing (8 byte object header):

patched Class uses 40 bytes less than original Class when empty
patched Class uses 208 bytes less than original Class when fully loaded
patched Class uses 16 bytes more than original Class when just one of 
fields is loaded

object instance counts:

patched Class uses the same number of object instances as original Class 
when empty
patched Class uses 6 object instances less than original Class when 
fully loaded
patched Class uses 1 object instance  more than original Class when just 
one of fields is loaded

The calculations are here:

The "worst case" scenario for patched code vs. original code is when 
just one of the fields in the structure is loaded. For example if only 
Class.getDeclaredMethods() or such is ever called on a Class instance. 
In all other cases (when Class is fresh or there are 2 or more fields 
loaded) the patched code is better.

The question remains how frequent the "worst case" scenario is in 
real-world code.

Regards, Peter

On 11/07/2012 11:39 AM, Peter Levart wrote:
> On 11/07/2012 03:10 AM, David Holmes wrote:
>> Hi Peter,
>> The movement of the reflection caches to a helper object is exactly 
>> what I had previously proposed here (some differences in the details 
>> of course):
>> and discussed here:
>> but this did not touch the annotations fields.
>> David
> Hi David,
> Thanks for the pointer. There is a discussion between Brian and you 
> (to quote some of it):
> On 5/04/2012 1:28 PM, Brian Goetz wrote:
> >/  Reducing the number of SoftReferences in ReflectionHelper also seems an
> />/  attractive target for memory reduction. Rather than eight soft
> />/  references (eight extra objects), maintaining a SoftRef to the entire
> />/  RH, or at least to the part of the RH that is currently SR'ed if the two
> />/  non-SR'ed fields can't be recomputed, would save you a whole pile of
> />/  objects per class (and might also reduce pressure on GC, having 8x fewer
> />/  SRs to process.)
> /
> I'd have to consider the intended semantics of these soft references
> before considering any change here. It would hard to predict how this
> might impact runtime performance if we have coarser-grained soft
> references. The current changes should be semantically null.
> >/  Finally, you may be able save an extra field per Class by storing the
> />/  ReflectionHelper in a ClassValue on Java SE 8, rather than a field.
> /
> ClassValue is something I'm keeping an eye on, but an entry in
> ClassValue is going to consume more dynamic memory than a single direct
> field.
> Thanks,
> David
> ...the 8 SoftReferences refer to arrays which are never hard 
> referenced by the outside world (arrays are always defensively 
> copied), so it's reasonable to expect that all SoftReferences would be 
> cleared at the same time anyway. And if 8 SoftReferences are replaced 
> with 1, the worst case scenario (to quote Hinkmond Wong):
> Hi Brian,
> One of the issues we have in the Java Embedded group (as David points
> out in his summary), is that while on paper the theoretical max savings
> seems great (as you point out also), in practice as David points out in
> his note, this might be a wash if there are a lot more reflection using
> classes vs. non-reflection using classes in "typical" real-world
> applications, not the low or zero reflection using class ratio that
> happens in the theoretical "best case".
> So, a question comes up if we should judge the merit of this change on
> the theoretical "best case" scenario, or should we judge it on
> real-world applicability to "typical" apps (such as a finite set of
> customer surveyed embedded apps that we feel represent a real-world
> scenario).
> Thanks,
> Hinkmond
> ...actually becomes even more favourable. We reduce huge overhead 
> (each SoftReference is 4 OOPs and 1 long). And if this single 
> SoftReference is ever cleared, more memory is released - the whole 
> structure (ReflectionHelper / VolatileData)
> Other differences in details between your proposal and mine:
> In your proposal, the method ReflectionHelper rh() is equivalent to 
> mine VolatileData volatileData() - it lazily constructs the structure 
> and returns it. My implementation also incorporates the logic of 
> clearCachesOnClassRedefinition() by returning and installing a new 
> instance of the structure in case a redefinition is detected. This has 
> a profound impact on the correctness of the cached data in face of 
> races that can occur.
> In your proposal, even if the VM could atomically publish changes to 
> raw reflection data and the classRedefinitionCount at the same time 
> (we hope that at least the order of publishing is such that 
> classRedefinitionCount is updated last), it can theoretically be that 
> 2 or more redefinitions of the same class happen in close proximity:
> VM thread: redefines the class to version=1
> thread 1: clears the cache and takes  version=1 raw data and computes 
> derived data but gets pre-empted before installing it
> VM thread: redefines the class to version=2
> thread 2: clears the cache and takes  version=2 raw data and computes 
> derived data and installs it
> thread 1: ...gets back and installs version=1 derived data over 
> version=2 data
> ...if there are no more class redefinitions, the stale version of 
> derived data can persist indefinitely.
> In my proposal, each thread will get it's own copy of the structure in 
> the above scenario and install the derived data into it. It can happen 
> that a particular instance of the structure does not represent a 
> "snapshot" view of the world, but that is not important, since that 
> particular inconsistent instance is only used for the in-flight call 
> and only in that part that is consistent. Other callers will get a 
> fresh instance.
> There is also one thing I overlooked and you haven't: the 
> cachedConstructor and newInstanceCallerCache fields.
> I'll have to look at how to incorporate them into my scheme. They are 
> currently neither SoftReferenced nor cleared at class redefinition. As 
> the cachedConstructor is only used to implement the .newInstance() 
> method, I wonder if it is safe not to clear it when the class is 
> redefined. Are old versions of Constructors still valid for invoking 
> in a redefined class? I guess they must be, since user code is free to 
> cache it's own versions and class redefinition should not prevent 
> invoking them...
> Since cachedConstructor/newInstanceCallerCache are used to optimize 
> .newInstance() method. That alone suggests that calling this method is 
> more common use-case than others. Perhaps leaving this pair of fields 
> out of the game is a better approach space-saving wise.
> Regards, Peter
>> On 6/11/2012 11:12 PM, Peter Levart wrote:
>>> On 11/05/2012 06:23 AM, David Holmes wrote:
>>>> Hi Peter,
>>>> Moving the annotations fields into a helper object would tie in with
>>>> the Class-instance size reduction effort that was investigated as part
>>>> of "JEP 149: Reduce Core-Library Memory Usage":
>>>> The investigations there to date only looked at relocating the
>>>> reflection related fields, though the JEP mentions the annotations as
>>>> well.
>>>> Any such effort requires extensive benchmarking and performance
>>>> analysis before being accepted though.
>>>> David
>>>> -----
>>> On 11/05/2012 10:25 AM, Alan Bateman wrote:
>>>> Here's a good starting place on the interaction of runtime visible
>>>> attributes and RedefineClasses/RetransformClasses:
>>>> -Alan.
>>> Hi all,
>>> Presented here is a patch mainly against java.lang.Class and also
>>> against java.lang.reflect.[Field,Method,Constructor,Executable] 
>>> classes.
>>> Currently java.lang.Class uses the following fields to maintain caches
>>> of reflection data that are invalidated as a result of class or
>>> superclass redefinition/re-transformation:
>>> private volatile transient SoftReference<Field[]> declaredFields;
>>> private volatile transient SoftReference<Field[]> publicFields;
>>> private volatile transient SoftReference<Method[]> declaredMethods;
>>> private volatile transient SoftReference<Method[]> publicMethods;
>>> private volatile transient SoftReference<Constructor<T>[]>
>>> declaredConstructors;
>>> private volatile transient SoftReference<Constructor<T>[]>
>>> publicConstructors;
>>> private volatile transient SoftReference<Field[]> declaredPublicFields;
>>> private volatile transient SoftReference<Method[]> 
>>> declaredPublicMethods;
>>> // Value of classRedefinedCount when we last cleared the cached values
>>> // that are sensitive to class redefinition.
>>> private volatile transient int lastRedefinedCount = 0;
>>> // Annotations cache
>>> private transient Map<Class<? extends Annotation>, Annotation> 
>>> annotations;
>>> private transient Map<Class<? extends Annotation>, Annotation>
>>> declaredAnnotations;
>>> If I understand Alan's references correctly, current VM can redefine 
>>> the
>>> class in a way that changes method bodies. Also new methods can be
>>> added. And the set of annotations can also be altered. And future
>>> improvements could allow even more.
>>> Because annotations are cached on Field/Method/Constructor instances,
>>> all the above fields must be invalidated when the class or 
>>> superclass is
>>> redefined.
>>> It can also be observed that Field/Method/Constructor caches are
>>> maintained using SoftReferences but annotations are hard references. I
>>> don't know if this is intentional. I believe that annotations could 
>>> also
>>> be SoftReferenced, so that in the event of memory pressure they get
>>> cleared. Many applications retrieve annotations only in the early 
>>> stages
>>> of their life-cycle and then either cache them themselves or forget
>>> about them.
>>> So I designed the patch to equalize this. If this is undesirable, the
>>> patch could be modified to make a distinction again.
>>> The patch replaces the above-mentioned java.lang.Class fields with a
>>> single field:
>>> private volatile transient SoftReference<VolatileData<T>> volatileData;
>>> ...which is a SoftReference to the following structure:
>>> // volatile data that might get invalid when JVM TI 
>>> RedefineClasses() is
>>> called
>>> static class VolatileData<T> {
>>> volatile Field[] declaredFields;
>>> volatile Field[] publicFields;
>>> volatile Method[] declaredMethods;
>>> volatile Method[] publicMethods;
>>> volatile Constructor<T>[] declaredConstructors;
>>> volatile Constructor<T>[] publicConstructors;
>>> // Intermediate results for getFields and getMethods
>>> volatile Field[] declaredPublicFields;
>>> volatile Method[] declaredPublicMethods;
>>> // Annotations
>>> volatile Map<Class<? extends Annotation>, Annotation> annotations;
>>> volatile Map<Class<? extends Annotation>, Annotation> 
>>> declaredAnnotations;
>>> // Value of classRedefinedCount when we created this VolatileData 
>>> instance
>>> final int redefinedCount;
>>> So let's look at static overhead differences using 64 bit addressing
>>> (useful load - arrays, Maps not counted since the patched code uses the
>>> same amount of same types of each).
>>> * Fresh java.lang.Class instance:
>>> current JDK8 code:
>>> 10 OOPs + 1 int = 10*8+4 = 84 bytes in 1 instance
>>> vs. patched code :
>>> 1 OOP = 8 bytes in 1 instance
>>> * Fully loaded java.lang.Class (Fields, Methods, Constructors, 
>>> annotations):
>>> current JDK8 code:
>>> 10 OOPs + 1 int = 84 bytes
>>> 8 SoftReference instances = 8*(header + 4 OOPs + 1 long) = 
>>> 8*(24+32+8) =
>>> 8*64 = 512 bytes
>>> total: 84+512 = 596 bytes in 9 instances
>>> vs. patched code :
>>> 1 OOP = 8 bytes
>>> 1 SoftReference = 64 bytes
>>> 1 VolatileData = header + 10 OOPs + 1 int = 24+84 = 108 bytes
>>> total: 8+64+108 = 180 bytes in 3 instances
>>> So we have 84 vs. 8 (empty) or 596 vs. 180 (fully loaded) byte 
>>> overheads and
>>> 1 vs. 1 (empty) or 9 vs. 3 (fully loaded) instance overheads
>>> Other than that, the patch also removes synchronized blocks for lazy
>>> initialization of annotations in Class, Field, Method and Constructor
>>> and replaces them with volatile fields. In case of Class.volatileData,
>>> this field is initialized using a CAS so there is no race which could
>>> install an already stale instance over more recent. Although such race
>>> would quickly be corrected at next call to any retrieval method, 
>>> because
>>> redefinedCount is now an integral part of the cached structure not an
>>> individual volatile field.
>>> There is also a change in how annotations are cached in Field, Method
>>> and Constructor. Originally they are cached in each copy of the
>>> Field/Method/Constructor that is returned to the outside world at each
>>> invocation of Class.getFields() etc. Such caching is not very effective
>>> if the annotations are only retrieved once per instance. The patch
>>> changes this and delegates caching to the "root" instance which is held
>>> inside Class so caching becomes more effective in certain usage
>>> patterns. There's now a possible hot-spot on the "root" instance but
>>> that seems not to be a bottleneck since the fast-path does not involve
>>> blocking synchronization (just volatile read). The effects of this
>>> change are clearly visible in one of the benchmarks.
>>> I have tried to create 3 micro benchmarks which exercise concurrent 
>>> load
>>> on 3 Class instances.
>>> Here's the benchmark code:
>>> And here are the results when run on an Intel i7 CPU (4 cores, 2
>>> threads/core) Linux machine using -Xmx4G VM option:
>>> The huge difference of Test1 results is a direct consequence of patched
>>> code delegating caching of annotations in Field/Method/Constructor to
>>> the "root" instance.
>>> Test2 results show no noticeable difference between original and 
>>> patched
>>> code. This, I believe, is the most common usage of the API, so another
>>> level of indirection does not appear to present any noticeable
>>> performance overhead.
>>> The Test3 on the other hand shows the synchronization overhead of
>>> current jdk8 code in comparison with non-blocking synchronization in
>>> patched code.
>>> JEP 149 also mentions testing with SPECjbb2005 and SPECjvm98, but that
>>> exceeds my possibilities.
>>> The patch against jdk8/jdk8/jdk hg repository is here:
>>> You can also browse the changed sources:
>>> Regards, Peter

More information about the core-libs-dev mailing list