harmony-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Evgueni Brevnov" <evgueni.brev...@gmail.com>
Subject Re: The way to increase the SPECjAppServer2004 score by 20%
Date Thu, 25 Oct 2007 07:18:04 GMT
This test shows two things:

1) SUN doesn't strictly follow the spec since it doesn't call
loadClass() method of the parent class loader but uses alternative
call chain instead.

2) Every class loader in the calling chain is marked as initiated
class loader for class hi since findLoadedClass() returns non null
value.

I believe if you change class loader's parent on the fly it will not
have any effect on already loaded classes. So it should be legal (and
compatible with SUN) to cache classes.

Thanks
Evgueni



On 10/24/07, Sergey Dmitriev <sergey.v.dmitriev@gmail.com> wrote:
> Hello
>
> sorry for such a delay. Here what I've got with my experiments. Please
> refer to  classloadertest2.java attached to JIRA.
>
> Talking about the Pavel's case: CL1 <- CL2 <- CL3, loading via CL3
> then loading via CL2.
>
> SUN JDK 6:
> mycl1.parent = sun.misc.Launcher$AppClassLoader@601bb1
> mycl2.parent = [MyClassLoader [dir:my]]
> mycl3.parent = [MyClassLoader [dir:X2]]
> [MyClassLoader [dir:X3]].loadClass(hi) enter...
> [MyClassLoader [dir:my]].findClass(hi) enter...
> [MyClassLoader [dir:my]].loadClassData(hi) enter...
> [MyClassLoader [dir:my]].loadClassData(hi) exit. out = [B@1e5e2c3
> [MyClassLoader [dir:my]].findClass(hi) exit. out = [B@1e5e2c3
> [MyClassLoader [dir:my]].loadClass(java.lang.Object) enter...
> [MyClassLoader [dir:my]].loadClass(java.lang.Object) exit. out = class
> java.lang.Object
> [MyClassLoader [dir:X3]].loadClass(hi) exit. out = class hi
> Class.forName("hi", true, mycl3) = [class hi]
> [MyClassLoader [dir:X2]].loadClass(hi) enter...
> [MyClassLoader [dir:X2]].loadClass(hi) exit. out = class hi
> Class.forName("hi", true, mycl2) = [class hi]
> mycl1.findLoadedClass(hi) = [class hi]
> mycl2.findLoadedClass(hi) = [class hi]
> mycl3.findLoadedClass(hi) = [class hi]
>
> HARMONY:
> mycl1.parent = java.lang.ClassLoader$SystemClassLoader@1089e19e
> mycl2.parent = [MyClassLoader [dir:my]]
> mycl3.parent = [MyClassLoader [dir:X2]]
> [MyClassLoader [dir:X3]].loadClass(hi) enter...
> [MyClassLoader [dir:X2]].loadClass(hi) enter...
> [MyClassLoader [dir:my]].loadClass(hi) enter...
> [MyClassLoader [dir:my]].findClass(hi) enter...
> [MyClassLoader [dir:my]].loadClassData(hi) enter...
> [MyClassLoader [dir:my]].loadClassData(hi) exit. out = [B@108d9a98
> [MyClassLoader [dir:my]].findClass(hi) exit. out = [B@108d9a98
> [MyClassLoader [dir:my]].loadClass(java.lang.Object) enter...
> [MyClassLoader [dir:my]].loadClass(java.lang.Object) exit. out = class
> java.lang.Object
> [MyClassLoader [dir:my]].loadClass(hi) exit. out = class hi
> [MyClassLoader [dir:X2]].loadClass(hi) exit. out = class hi
> [MyClassLoader [dir:X3]].loadClass(hi) exit. out = class hi
> Class.forName("hi", true, mycl3) = [class hi]
> [MyClassLoader [dir:X2]].loadClass(hi) enter...
> [MyClassLoader [dir:my]].loadClass(hi) enter...
> [MyClassLoader [dir:my]].loadClass(hi) exit. out = class hi
> [MyClassLoader [dir:X2]].loadClass(hi) exit. out = class hi
> Class.forName("hi", true, mycl2) = [class hi]
> mycl1.findLoadedClass(hi) = [class hi]
> mycl2.findLoadedClass(hi) = [null]
> mycl3.findLoadedClass(hi) = [null]
>
> First thing that one can catch: in case of SUN we don't touch the CL2
> at all during the loading via CL3. Interesting.
>
> And during the loading via CL2 SUN does not touch the CL1.
>
>
> Thanks
> Sergey
>
> On 10/23/07, Evgueni Brevnov <evgueni.brevnov@gmail.com> wrote:
> > Pavel,
> >
> > I don't know exact answer to you question :-(, sorry, I hope Dmitry
> > will provide such information soon :-)
> >
> > I have two notes:
> >
> > 1) I believe the spec requires to return the same class each time
> > since a pair <ClassLoader, ClassName> must uniquely identify java
> > class. So in you example from the first letter CL2 should always
> > return the same class.
> >
> > 2) Regarding "initiating" class loader notion. My understanding is
> > that each class loader in the hierarhy chain up to defining one should
> > be marked as initiating class loader. At least this is written in  [1]
> >
> > [1] "Inside the Jva2 Virtual Machine" by Bill Venners.
> >
> > Thanks
> > Evgueni
> > On 10/23/07, Pavel Pervov <pmcfirst@gmail.com> wrote:
> > > Evgueni,
> > >
> > > 1) Our own URLClassLoader overrides ClassLoader.loadClass method, so the
> > > idea of adding such registry to base class will not work from the start.
> > >
> > > 2) The question in my previous letter is not answered. If the answer is
> > > positive, then adding the registry of "initiated" classes to URLClassLoader
> > > (which AFAIU will cover the issue described by Sergey) still requires solid
> > > algorithm to distinguish the class loader which was initially asked to load
> > > a class from the one, which resides in the middle of delegation chain.
> > >
> > > I'm not against caching, I just see problems I do not know how to solve.
> > >
> > > WBR,
> > >    Pavel.
> > >
> > > P.S. BTW, we have the registry of "initiated" classes but it is only
> > > appended on class resolutions.
> > > On 10/23/07, Evgueni Brevnov <evgueni.brevnov@gmail.com> wrote:
> > > >
> > > > Pavel,
> > > >
> > > > As you pointed out to change delegation model user must override
> > > > ClassLoader.loadClass() method. By doing so user can specify any
> > > > search order he only want and doesn't depend on caching capabilities
> > > > provided by java.lang.ClassLoader. If user uses default implementation
> > > > of ClassLoader.loadClass() method then it should be safe to use cache.
> > > > What do you think?
> > > >
> > > > Thanks
> > > > Evgueni
> > > >
> > > > On 10/23/07, Pavel Pervov <pmcfirst@gmail.com> wrote:
> > > > > Evgueni,
> > > > >
> > > > > ClassLoader.loadClass is not final. One can overload this and build
any
> > > > > delegation model his or her application requires. So, -1 here.
> > > > >
> > > > > It would be good to answer the following case before considering
class
> > > > > caching.
> > > > >
> > > > > Assume, we have delegation of three class loaders: CL1<-CL2<-CL3.
> > > > > Assume, class Foo is only available from CL1.
> > > > >
> > > > > We call to CL3.loadClass("Foo"). The call chain will look like
> > > > > CL3.loadClass->CL2.loadClass->CL1.loadClass. As far as we know,
the
> > > > second
> > > > > and subsequent calls to CL3.loadClass("Foo") will not call to
> > > > > CL2.loadClass("Foo")
> > > > > but return "cached" copy from CL3 on RI (as described by Sergey).
> > > > >
> > > > > But it would be really interesting to know if the call to CL2.loadClasswill
> > > > > result in call chain CL2.loadClass->CL1.loadClass.
> > > > >
> > > > > If this is true, my initial idea that it is only allowed to register
the
> > > > > class in "initiating" and "defining" loaders is also true. And I
have no
> > > > > solid solution to determining "initiating" class loader.
> > > > >
> > > > > Otherwise, it is ok to create the cache in URLClassLoader as Sergey
> > > > > initially suggested.
> > > > >
> > > > > Pavel.
> > > > >
> > > > > On 10/23/07, Evgueni Brevnov <evgueni.brevnov@gmail.com> wrote:
> > > > > >
> > > > > > Pavel,
> > > > > >
> > > > > > I think its not common situation when parent class loader is
changed
> > > > > > at runtime. To be honest I don't know how it's possible since
parent
> > > > > > class loader is kept in private field. Anyway, can we somehow
detect
> > > > > > such situation and switch to normal scheme or discard cached
classes
> > > > > > upon switching parent class loader.
> > > > > >
> > > > > > Thanks
> > > > > > Evgueni
> > > > > >
> > > > > > On 10/22/07, Pavel Pervov <pmcfirst@gmail.com> wrote:
> > > > > > > Looking at the Java Virtual Machine Specification there
are two
> > > > terms:
> > > > > > > initiating loader and defining loader.
> > > > > > >
> > > > > > > Initiating loader is the class loader which was first asked
to load
> > > > > > class in
> > > > > > > delegation chain (CL2 in your example).
> > > > > > >
> > > > > > > Defining loader is the class loader which has found bytes
for class
> > > > > > > definition and called to JNI->DefineClass.
> > > > > > >
> > > > > > > Lets review the following delegation model.
> > > > > > >
> > > > > > > CL1<-CL2<-CL3
> > > > > > >
> > > > > > > CL4<-CL2<-CL5
> > > > > > >
> > > > > > > There is nothing wrong with CL2: its delegation model simply
assumes
> > > > > > > switching delegation parent depending on some external
property.
> > > > > > >
> > > > > > > Different versions of class Foo are available from CL1
and CL4. Now,
> > > > we
> > > > > > do
> > > > > > > something like CL3.loadClass("Foo") and CL5.loadClass("Foo").
> > > > > > > If CL2 will cache the class Foo, then the call to CL5 will
return
> > > > wrong
> > > > > > > result.
> > > > > > >
> > > > > > > So the caching is possible, but it is only allowed to register
class
> > > > in
> > > > > > > initiating loader and in defining loader. Otherwise, you
may have
> > > > wrong
> > > > > > > classes returned to your application.
> > > > > > >
> > > > > > > Determining defining loader is simple. Can you suggest
some working
> > > > idea
> > > > > > on
> > > > > > > how to determine initiating loader for a class?
> > > > > > >
> > > > > > > WBR,
> > > > > > >    Pavel.
> > > > > > >
> > > > > > > On 10/22/07, Sergey Dmitriev <sergey.v.dmitriev@gmail.com>
wrote:
> > > > > > > >
> > > > > > > > Hello
> > > > > > > >
> > > > > > > > This letter is actually about HARMONY-4965, I just
changed a
> > > > little
> > > > > > > > bit a subject. :)
> > > > > > > >
> > > > > > > > During my play with SPECjAppServer2004 on OC4J I've
found out that
> > > > we
> > > > > > > > have a way to optimize the class loader. Or in other
words
> > > > currently
> > > > > > > >
> > > > > > > > Lets say we have class loaders CL1 and CL2, and CL2.getParent()
==
> > > > > > > > CL1. And lets say we have a class C which is loaded
by the CL1
> > > > (the
> > > > > > > > parent one). And then we try to load class via the
class loader
> > > > CL2:
> > > > > > > > of course CL2 first of all tries to load it via its
parent CL1 and
> > > > > > > > succeed in it. When we try to load class C via CL2
again we will
> > > > have
> > > > > > > > the same chain of calls: CL2 -> CL1. But if CL2
have support of
> > > > > > > > caching of classes loaded by not itself - the chain
can become
> > > > much
> > > > > > > > shorter - it could end at CL2.findFinishedClass().
> > > > > > > >
> > > > > > > > For example RI behaves like this. See JIRA's demo.
> > > > > > > >
> > > > > > > > Back to the SPECjAppServer2004 @ OC4J. In this case
we have a
> > > > deeper
> > > > > > > > chain: we have a concrete class which is been searched
so deep
> > > > that we
> > > > > > > > looking for this class in boot strap class path. And
this is what
> > > > we
> > > > > > > > have in steady state of the benchmark. Don't you think
it is
> > > > useless?
> > > > > > > > :)
> > > > > > > >
> > > > > > > > Talking about numbers: my measurements show that we
can take up to
> > > > 20%
> > > > > > > > of burst having this bug fixed. Could some competent
guy take a
> > > > look
> > > > > > > > into this?
> > > > > > > >
> > > > > > > >
> > > > > > > > Thanks
> > > > > > > > Sergey
> > > > > > > >
> > > > > > >
> > > > > > >
> > > > > > >
> > > > > > > --
> > > > > > > Pavel Pervov,
> > > > > > > Intel Enterprise Solutions Software Division
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > > >
> > > > > --
> > > > > Pavel Pervov,
> > > > > Intel Enterprise Solutions Software Division
> > > > >
> > > >
> > >
> > >
> > >
> > > --
> > > Pavel Pervov,
> > > Intel Enterprise Solutions Software Division
> > >
> >
>

Mime
View raw message