Sorry I must correct myself... the test are making 1000000 for the free EntityModes so this mean that the old implementation work in 1 second per 1000000 of persistent classes.BTW patch applied and issue closed.
The end of this story sound like the worry about DynProxy getter/setter 2009/2/7 Fabio Maulo <[email protected]> > BTW, to understand how much important it is, without the apply the patch > and iteration of 1000000 take around 1 second.1000000, at most divided by > 3, mean that you are working with 333333 persistent classes. > > 2009/2/7 Fabio Maulo <[email protected]> > > The problem is the boxing in GetHashCode method of the enumerator. >> >> >> 2009/2/6 Roger Kratz <[email protected]> >> >>> Yes, it's put on the JIRA. There's a impl of >>> IEqualityComparer<EntityMode> and "hard coded" in the constructor for the >>> LinkedHashMap in EntityModeToTuplizerMapping. That's why a factory might be >>> better (both used for dictionary and linkedhashmap and (?) some other >>> IDictionary impl)? >>> >>> >>> >>> *From:* [email protected] [mailto: >>> [email protected]] *On Behalf Of *Fabio Maulo >>> *Sent:* den 6 februari 2009 16:57 >>> >>> *To:* [email protected] >>> *Subject:* [nhibernate-development] Re: IDictionary<entitymode, xxx> >>> >>> >>> >>> Roger, >>> >>> BTW the EntityModeEqualityComparer as separated class is need (when we >>> must use some other implementation of IDictionary). >>> >>> 2009/2/6 Roger Kratz <[email protected]> >>> >>> ...or... >>> >>> New EntityModeDic<tvalue>() >>> >>> ...or... >>> >>> EntityMode.CreateDic<tvalue() >>> >>> …or… >>> >>> ?? >>> >>> >>> >>> >>> >>> *From:* [email protected] [mailto: >>> [email protected]] *On Behalf Of *Fabio Maulo >>> *Sent:* den 6 februari 2009 16:51 >>> >>> >>> *To:* [email protected] >>> *Subject:* [nhibernate-development] Re: IDictionary<entitymode, xxx> >>> >>> >>> >>> Well... this is "our" work around; the real issue should be send to >>> Microsoft. >>> >>> I hope Microsoft solve it soon; the hope is the last to die ;) >>> >>> >>> >>> In NH we have few IDictionary<K,V> depending on enum and probably all are >>> using EntityMode as K so, probably, an implementation >>> of EntityModeEqualityComparer and a simple modification to the Dictionary >>> constructor call is enough. >>> >>> 2009/2/6 Roger Kratz <[email protected]> >>> >>> It's been added to the JIRA. >>> >>> >>> >>> However – if you think this should be handled in more places in the code, >>> I guess it could be more generic than this "isolated" patch. >>> >>> >>> >>> *From:* [email protected] [mailto: >>> [email protected]] *On Behalf Of *Fabio Maulo >>> *Sent:* den 6 februari 2009 16:34 >>> >>> >>> *To:* [email protected] >>> *Subject:* [nhibernate-development] Re: IDictionary<entitymode, xxx> >>> >>> >>> >>> If the performance issue is caused by EqualityComparer I prefer the >>> implementation of a specific EqualityComparer instead other solution. >>> >>> 2009/2/6 Roger Kratz <[email protected]> >>> >>> I can do a impl of the equalitycomparer and use that instead. >>> >>> >>> >>> Did a quick search in the source code and it seems to be other places >>> where dics are keyed by entitymode. I'm not sure which of them are used >>> heavily though, but I guess it wouldn't hurt to pass an IEqualityComparer to >>> them all? >>> >>> >>> >>> Should I do this? >>> >>> >>> >>> *From:* [email protected] [mailto: >>> [email protected]] *On Behalf Of *Ayende Rahien >>> *Sent:* den 6 februari 2009 16:01 >>> >>> >>> *To:* [email protected] >>> *Subject:* [nhibernate-development] Re: IDictionary<entitymode, xxx> >>> >>> >>> >>> Duh! >>> The implementation makes it obvious what is actually going on. >>> I think that a better alternative than int keys would be to use a >>> IEqualityComparer<TEnum> instead, that would keep the clarity, reduce >>> casting and prevent boxing. >>> >>> On Fri, Feb 6, 2009 at 4:53 PM, Roger Kratz <[email protected]> >>> wrote: >>> >>> That you'll better ask the dictionary implementors at Redmond J. I was a >>> bit surprised by this myself some time ago. Just create a little test >>> yourself and probably you will be more than surprised by the result… >>> >>> >>> >>> There are some articles out there describing it better than I probably >>> can… >>> >>> http://www.safnet.com/writing/tech/archives/2007/07/performance_3_c.html >>> >>> >>> http://blogs.msdn.com/shawnhar/archive/2007/07/02/twin-paths-to-garbage-collector-nirvana.aspx >>> >>> http://beardseye.blogspot.com/2007/08/nuts-enum-conundrum.html >>> >>> >>> >>> >>> >>> >>> >>> *From:* [email protected] [mailto: >>> [email protected]] *On Behalf Of *Ayende Rahien >>> *Sent:* den 6 februari 2009 15:45 >>> >>> >>> *To:* [email protected] >>> *Subject:* [nhibernate-development] Re: IDictionary<entitymode, xxx> >>> >>> >>> >>> Ha? >>> >>> >>> Why would there be boxing involved at all? >>> >>> On Fri, Feb 6, 2009 at 3:46 PM, Roger <[email protected]> wrote: >>> >>> >>> Hi >>> >>> EntityModeToTuplizerMapping holds a dic keyed by entitymode >>> internally. From a perf perspective, it's a bad thing to do – enum as >>> a key will cause a lot of boxing internally in dictionary operations. >>> And in (large) resultsets these dictionary items in it will be >>> retrieved _a lot_. >>> >>> I just made a quick test and changed this dictionary to be keyed by an >>> int instead (casting the enum to int when added), some extreme queries >>> (=fast query, big resultset) run approx 10% faster in total. >>> >>> I'll be glad to send this little patch. However, I would say a better >>> solution would be to change this enum to a class instead (I don't >>> really like enums at all :)). I can provide a patch for this as well, >>> but this type has a lot of dependencies. In short words – before >>> spending some hours doing this, I would like to see if you have any >>> opinions on this? >>> >>> Whatever you think, I would strongly suggest that get rid of often >>> used dictionaries keyed by enums. >>> >>> Keep up the great work! >>> >>> Regards >>> Roger >>> >>> >>> >>> >>> >>> >>> >>> >>> -- >>> Fabio Maulo >>> >>> >>> >>> >>> -- >>> Fabio Maulo >>> >>> >>> >>> >>> -- >>> Fabio Maulo >>> >> >> >> >> -- >> Fabio Maulo >> > > > > -- > Fabio Maulo > -- Fabio Maulo
