openjpa-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Kevin Sutter (JIRA)" <j...@apache.org>
Subject [jira] Commented: (OPENJPA-141) More performance improvements (in response to changes for OPENJPA-138)
Date Wed, 14 Feb 2007 23:11:06 GMT

    [ https://issues.apache.org/jira/browse/OPENJPA-141?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12473243
] 

Kevin Sutter commented on OPENJPA-141:
--------------------------------------

Yes, good catch, Craig.  The original location that I wanted to resolve the isAssignableFrom()
overhead was in BrokerImpl and that was just a one-way check.  The two-way check in FetchConfigurationImpl
was overlooked.  Thank you.  But, that brings up a new question...  Do we do the two-way check
in this new utility method (even though BrokerImpl didn't require this in the past)?  Or,
is the one-way check sufficient for FetchConfigurationImpl's usage?  Any historical perspective?

>>> 1. Why not keep a single assignable types map in ImplHelper?
>> I actually thought about doing that, but then I had concerns about the size of the
Cache.
> How are two static maps going to end up being smaller overall than one combined static
map?

Never mind.  Since we're dealing with (hopefully) unique Class keys, a single map will suffice.

>> Hmmm.. If I understand what you are saying, it really doesn't matter whether we use
hard, weak, or soft keys, since the resulting cache will be hard no matter what -- since we're
using Class objects in the cache.
> No, it does matter. And the type of value references we use matters way more. If we want
a "hard cache that drops entries for classes that are redeployed", then we should be using
weak keys and hard values. If we want a "memory sensitive" cache, then we should be using
hard keys and soft values. I'm not sure where the disconnect is coming from with these reference
types.

There have been several viewpoints on the use of these reference types and what the impact
would be.  To be honest, at this point, all that I am looking for is the ability to cache
these assignable types.  Whether it's redployment-friendly or memory-friendly, I don't really
care at this point.  We can worry about that later.  If you have a preference for this first
iteration, let me know.  Thanks.

>> The benchmark is a set of primitives based on the SpecJApp application using the
SunOne Application Server. The profiling data from this set of tests indicate that caching
of the JNDI lookup is beneficial.
>Beneficial over the suggested use of an instance variable? Or beneficial over no caching
of the TM whatsoever? There's a big difference.

Definitely beneficial over no caching of the TM whatsoever.  Sorry for the confusion.

>>> 5. Even if ImplHelper.isAssignable retains its map parameter (and per #1 above
I question why it should), it should just be a Map; I don't see why you'd have the method
require a ConcurrentMap.
>> I did this way to be thread safe. If I only used a Map parameter, then the caller
would have to ensure that any updates to the Cache are thread safe.
>The caller is giving you the Map in your scheme. It's up to him whether the Map he's giving
you is used concurrently or not. The helper method itself has no threading issues at all,
and only requires a Map. But I agree that if we move to a single cache in ImplHelper it's
a moot point.

That's definitely one way around it.  I prefer to enforce the requirement via the signature
of the contract.  Changing to a single map anyway...

> More performance improvements (in response to changes for OPENJPA-138)
> ----------------------------------------------------------------------
>
>                 Key: OPENJPA-141
>                 URL: https://issues.apache.org/jira/browse/OPENJPA-141
>             Project: OpenJPA
>          Issue Type: Sub-task
>          Components: jpa
>            Reporter: Kevin Sutter
>         Assigned To: Kevin Sutter
>         Attachments: openjpa-141.txt
>
>
> Abe's response to my committed changes for OPENJPA-138.  I will be working with Abe and
my performance team to work through these issues...
> > ======================================================================
> > ========
> > --- incubator/openjpa/trunk/openjpa-kernel/src/main/java/org/apache/
> > openjpa/ee/JNDIManagedRuntime.java (original)
> > +++ incubator/openjpa/trunk/openjpa-kernel/src/main/java/org/apache/
> > openjpa/ee/JNDIManagedRuntime.java Sun Feb 11 18:33:05 2007
> > @@ -29,6 +29,7 @@
> >      implements ManagedRuntime {
> >
> >      private String _tmLoc = "java:/TransactionManager";
> > +    private static TransactionManager _tm;
> Whoa, I didn't think you meant caching the TM statically.  That has
> to be backed out.  You can cache it in an instance variable, but not
> statically.  Nothing should prevent someone having two different
> BrokerFactories accessing two different TMs at two different JNDI
> locations.
> BrokerImpl:
> > +     * Cache from/to assignments to avoid Class.isAssignableFrom
> > overhead
> > +     * @param from the target Class
> > +     * @param to the Class to test
> > +     * @return true if the "to" class could be assigned to "from"
> > class
> > +     */
> > +    private boolean isAssignable(Class from, Class to) {
> > +      boolean isAssignable;
> > +      ConcurrentReferenceHashMap assignableTo =
> > +          (ConcurrentReferenceHashMap) _assignableTypes.get(from);
> > +
> > +      if (assignableTo != null) { // "to" cache exists...
> > +          isAssignable = (assignableTo.get(to) != null);
> > +          if (!isAssignable) { // not in the map yet...
> > +              isAssignable = from.isAssignableFrom(to);
> > +              if (isAssignable) {
> > +                  assignableTo.put(to, new Object());
> > +              }
> > +          }
> > +      } else { // no "to" cache yet...
> > +          isAssignable = from.isAssignableFrom(to);
> > +          if (isAssignable) {
> > +              assignableTo = new ConcurrentReferenceHashMap(
> > +                      ReferenceMap.HARD, ReferenceMap.WEAK);
> > +              _assignableTypes.put(from, assignableTo);
> > +              assignableTo.put(to, new Object());
> > +          }
> > +      }
> > +      return isAssignable;
> > +    }
> This code could be simplified a lot.  Also, I don't understand what
> you're trying to do from a memory management perspective.  For the
> _assignableTypes member you've got the Class keys using hard refs and
> the Map values using weak refs.  No outside code references the Map
> values, so all entries should be eligible for GC pretty much
> immediately.  The way reference hash maps work prevents them from
> expunging stale entries except on mutators, but still... every time a
> new entry is added, all the old entries should be getting GC'd and
> removed.  Same for the individual Map values, which again map a hard
> class ref to an unreferenced object value with a weak ref.  Basically
> the whole map-of-maps system should never contain more than one entry
> total after a GC run and a mutation.
> I'd really like to see you run your tests under a different JVM,
> because it seems to me like (a) this shouldn't be necessary in the
> first place, and (b) if this is working, it's again only because of
> some JVM particulars or GC timing particulars or testing particulars
> (I've seen profilers skew results in random ways like this) or even a
> bug in ConcurrentReferenceHashMap.
> The same goes for all the repeat logic in FetchConfigurationImpl.
> And if we keep this code or some variant of it, I strongly suggest
> moving it to a common place like ImplHelper.
> > +    /**
> > +     * Generate the hashcode for this Id.  Cache the type's
> > generated hashcode
> > +     * so that it doesn't have to be generated each time.
> > +     */
> >      public int hashCode() {
> >          if (_typeHash == 0) {
> > -            Class base = type;
> > -            while (base.getSuperclass() != null
> > -                && base.getSuperclass() != Object.class)
> > -                base = base.getSuperclass();
> > -            _typeHash = base.hashCode();
> > +            Integer typeHashInt = (Integer) _typeCache.get(type);
> > +            if (typeHashInt == null) {
> > +                Class base = type;
> > +                Class superclass = base.getSuperclass();
> > +                while (superclass != null && superclass !=
> > Object.class) {
> > +                    base = base.getSuperclass();
> > +                    superclass = base.getSuperclass();
> > +                }
> > +                _typeHash = base.hashCode();
> > +                _typeCache.put(type, new Integer(_typeHash));
> > +            } else {
> > +                _typeHash = typeHashInt.intValue();
> > +            }
> >          }
> >          return _typeHash ^ idHash();
> >      }
> Once again, you're mapping a hard Class ref to a value with no
> outside references held in a weak ref.  Once again that means the
> entry should be immediately eligible for GC, and therefore should be
> removed on the next mutation of the cache, subject to GC timing.  And
> again I'd like to know what your JVM is doing to make Class.hashCode
> take an appreciable amount of time.  Aren't Class instances supposed
> to be singletons?  What if we just used System.identityHashCode(cls)?
> > Modified: incubator/openjpa/trunk/openjpa-lib/src/main/java/org/
> > apache/openjpa/lib/conf/ObjectValue.java
> > URL: http://svn.apache.org/viewvc/incubator/openjpa/trunk/openjpa-
> > lib/src/main/java/org/apache/openjpa/lib/conf/ObjectValue.java?
> > view=diff&rev=506230&r1=506229&r2=506230
> > ======================================================================
> > ========
> > --- incubator/openjpa/trunk/openjpa-lib/src/main/java/org/apache/
> > openjpa/lib/conf/ObjectValue.java (original)
> > +++ incubator/openjpa/trunk/openjpa-lib/src/main/java/org/apache/
> > openjpa/lib/conf/ObjectValue.java Sun Feb 11 18:33:05 2007
> > @@ -17,6 +17,8 @@
> >
> >  import org.apache.commons.lang.ObjectUtils;
> >  import org.apache.openjpa.lib.util.Localizer;
> > +import org.apache.openjpa.lib.util.ReferenceMap;
> > +import
> > org.apache.openjpa.lib.util.concurrent.ConcurrentReferenceHashMap;
> >
> >  /**
> >   * An object {@link Value}.
> > @@ -28,6 +30,10 @@
> >      private static final Localizer _loc = Localizer.forPackage
> >          (ObjectValue.class);
> >
> > +    // cache the types' classloader
> > +    private static ConcurrentReferenceHashMap _classloaderCache =
> > +        new ConcurrentReferenceHashMap(ReferenceMap.HARD,
> > ReferenceMap.WEAK);
> This maps a hard Class ref to a weak ClassLoader ref.  Given that a
> Class references its ClassLoader (or is supposed to -- again I wonder
> what the hell the JVM you're using is doing where
> Class.getClassLoader is taking a long time), no entries will ever
> expire from this map.
> Have you tried running your benchmarks without all the caching of
> assignables and classloaders and hashcodes (all Class methods, btw)
> and just the other improvements?  Or with any other JVM?

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


Mime
View raw message